Digital Twins in Order Picking Systems for Operational Decision Support

Dirk KaukeTechnical University of Munich

dirk.kauke@tum.de

Stefan GalkaEast Bavarian Technical University

of Applied Sciences Regensburgstefan.galka@oth-regensburg.de

Johannes FottnerTechnical University of Munich

j.fottner@tum.de

Abstract

Digital twins are arousing great interest in bothscience and industry. There are a large number ofpapers that demonstrate and evaluate the potential ofDigital Twins in different application areas. However,it must be noted that there is still no uniform definitionof Digital Twins. This paper first examines the conceptof Digital Twins and highlight how they differ in level,compared with other digital models. The focus of thispaper lies in the conceptual development of a digitaltwin in order picking systems. The described approachin the paper aims at supporting the operational controlin order picking systems. Both the architecturalstructure and the functions, e.g. the simulation, aredescribed in detail. Overall, this thesis shows thebenefits of Digital Twins. However, some functionalextensions are still needed before the full potential canbe achieved.

1. Introduction

The term Digital Twin is used in many different waysin these days. In most cases, this means the virtualrepresentation of a product, process or system. The termDigital Twin was first used by Dr. Michael Grieves at theUniversity of Michigan, to explain a digital model as theequivalent of a real, physical object [1, 2, 3].

The research institute Gartner assumes that half of alllarge industrial companies will use a form of a digitaltwin by the year 2021 and that this will increase theireffectiveness by 10 percent [4].

The published digital twin examples differsignificantly in terms of use cases and technicalimplementation. Kritzinger et al. classify the differentapproaches into digital models, digital shadows andDigital Twins (shown in Fig. 1) [5]. The essentialfeature of the Digital Twins is the bidirectional dataexchange between the real system and the digitalrepresentation [6, 5, 3, 2]. The frequency of dataexchange between the real and the virtual systems is not

specified.Due to the data exchange in both directions, Digital

Twins are suitable for integration into decision-supportsystems, because they represent the current situation inthe system (database), and feasibility of decisions canbe checked in advance using the virtual system. Thisreduces the risk of making an incorrect decision.n termsof decision support and the review of any decisions,future requirements must be considered; thereforeit makes sense to use Digital Twins together withforecasting models. In the context of data analytics, thisis described as predictive or prescriptive analytics.

A purposeful and forward-looking optimization ofcomplex logistics and production systems during theoperation phase is only possible with appropriate tools.A digital twin of the system expanded by functionsof predictive / prescriptive analysis is a promisingapproach. Typical areas of application are productioncontrol, the adjustment of machine parameters, andemployee resource planning [5].

Figure 1 illustrates the correlation between thedifferent levels of digital models and data analytics. Thefigure also shows which data basis is generally used andwho makes the decisions.

In data harvesting, data is collected and analyzedfor a system. There is no intention to model systems,as occurs in the concepts of the digital model, digitalshadows and digital twins. Data harvesting is often usedin combination with the other approaches.

Digital models (level 1) support the creation ofreports and the retrospective / descriptive analysis ofsystems. Due to the lack of automation in the dataexchange, these analyses are triggered in larger intervalsand for specified questions (used in strategic planning).

Digital shadows (level 2) allow automatic datatransfer from real to virtual systems. This means thata large amount of data is available for analysis. Themodeling of system behavior (interdependencies) is lesspronounced than in Digital Twins. For this reason,digital shadows are well suited for the analysis ofindividual areas in a system (e.g. one machine). The

Proceedings of the 54th Hawaii International Conference on System Sciences | 2021

Page 1655URI: https://hdl.handle.net/10125/70812978-0-9981331-4-0(CC BY-NC-ND 4.0)

Figure 1. Different levels of digital models. Authors’ own representation based on [5]

large database allows forecasts e.g. for maintenanceneeds. Digital twins (level 3) also have a largedatabase; in addition, the interactions between thesystem elements are modeled, and this allows the systembehavior to be described. For this purpose, DigitalTwins often have a simulation model for which therequired data is automatically processed. Digital twinstherefore support the prescriptive analytics approach.

The TwinKom project, which is presented in thispaper, introduces a digital twin, uses approachesof predictive and prescriptive analytics, and makesdecision-suggestions to a user or takes decisions on itsown.

2. Digital Twins in logistics

In the previous chapter, it has already been pointedout that there is no standard definition of a digital twinas yet. This also becomes clear when logistics areconsidered as a field of application. Exemplary studies,on closer examination, tend to focus on a digital model(Level 1) or model (Level 2) [7, 8, 9, 10, 11].

A detailed literature review was carried out to revealtwo scientific papers that contain descriptions of adigital twin in logistics [12, 13]. Braglia et al. havedesigned a digital twin that supports forklift routing. Byusing an unspecified data interface, the position data ofthe forklift and the pallets to be transported are supposed

to be transmitted to the digital model in a definedtime interval. The simulation model examines variousroute and order strategies based on the available data.Afterwards, the respective optimal route is transmittedto the forklift driver [12].

Korth et al. also present in their work a conceptfor a digital twin for the control of logistics systems.In contrast to Braglia et al., they specify the datainterfaces and try to collect data through scanning orsensors. However, a more detailed procedure for theconcrete implementation of the information flow cannotbe inferred from the paper. As an application examplefor the digital twin, a distribution center consisting of anautomatic pallet shelf warehouse, two picking areas andinbound and outbound goods departments is presented.The objective in this case is the balanced utilizationof the employees in the outgoing goods department.The presented result data could be achieved. If theapplicability of the Digital Twins, even in the muchmore complex order picking system, is in the paper notmore closely examined [13]. The papers shown confirmthat Digital Twins are often used for operational issues.

Page 1656

Table 1. Overview of literature regarding digital twins and related concepts

Digital Shadow Digital Twin Internal Logistics External Logistics

Ashrafian et al. (2019) X XBaruffaldi et al. (2018) X XBraglia et al. (2019) X XBuckova et al. (2019) X XHofmann & Branding (2019) X XKorth et al. (2018) X XLeng et al. (2019) X X

3. Digital Twins for order picking systems

3.1. Related literature of order pickingsystems

Order picking has become one of the most importantprocesses in intralogistics during the development ofonline commerce [14]. The storage of a large numberof different articles as well as the immediate deliveryof the goods are significantly influenced by the orderpicking system. Therefore, the optimal design of thepicking system has become a significant competitiveadvantage, especially for many online retailers. Theoperation of the system is aligned with the followingfour objectives.

• High performance

• Low costs

• Short throughput times

• High adherence to delivery dates

The performance of the picking system, measuredin picks/h, ensures that the customer orders can beprocessed in time. At the same time, the targetperformance should be achieved at the lowest possiblecost, to ensure that the system is operated as efficientlyas possible. Short lead times have a decisive influenceon the overall performance. In addition, buffer stocksshould be kept as low as possible in multi-area systems.Cut-off times should ultimately ensure that deadlines aremet.

The achievement of these goals has been the subjectof scientific investigation for decades. This hasled to the development of various strategies which,depending on their characteristics, can influence thesystem positively or negatively. In the course of time,several comprehensive literature reviews have beenpublished, to which reference should be made here

[15, 16]. For the development of this scientific work thework of Van Gils et. al. will be primarily considered.

The parameters to be examined are based on ageneral division into strategic, tactical and operationalplanning problems [16]. Strategic parameters includelayout design, degree of automation and handlingequipment. The digital twin is not supposed toinvestigate any variation of these parameters, althoughit is more likely that an overall system can consistof different subsystems with different degrees ofautomation, layouts and handling equipment. Theoperational parameters are assigned to the resourcedimension (to which the number of people in the systemalso belongs), zoning and storage assignment. Batching,routing and job assignment, among other things, areconsidered in operational terms.

In the scope of this classification, no time intervalwas considered in which the planning problems areexamined or changed. There are reasonable groundsto assume that the frequency will increase significantlyfrom strategic to operational. Since the digital twinis supposed to consider parameters at intervals rangingfrom daily to hourly or shorter, the tactical andoperational parameters are, in principle, relevant.

In order to achieve these goals, however, there isscope for action in the operation of heterogeneous orderpicking systems (multi-area system), which makes itpossible to react to changing conditions. The followingchapter will present the scope and potential of DigitalTwins in order picking systems in more detail.

3.2. Application and potentials

In order to continuously ensure the definedobjectives of a picking system, a regular review ofcertain actions is necessary. The scope for actionconsists of three key aspects: The size of picking toursis one of the principal levers. A picking tour can existof one or more (multi-order-picking) orders. The task ofthe Digital Twin is to find the optimal size of the tour sothat the requirements for lead times and efficiency can

Page 1657

be met.A dynamic human-resource-planning system allows

the flexible distribution of available employees to thedifferent areas of the picking system according to thecurrent order situation. In this way, the daily fluctuatingorders can be counteracted and a high utilization of theworkforce can be achieved.

In addition, a time-differentiated release of pickingorders makes it possible to synchronize the expectedarrival time of partial orders in consolidation, so thatbuffer areas can be reduced and the reliability of thepicking system can be increased.

These actions are powerful tools for overcomingthe existing challenges, but these days they are mainlybased on human experience and controlled by simplekey figures. It is often not possible to fully evaluatethe consequences of individual decisions. A digitaltwin, on the other hand, should be able to select apackage of measures from the room for maneuver, bymonitoring, forecasting and simulating in such a waythat the requirements are met as well as possible.

3.3. Research methology

The digital twin consists essentially of a simulationmodel. Around the core, different services (functions)are arranged, which are described in the systemsarchitecture. The procedure of this research is thereforebased on the usual procedure of simulation projects,which starts with a problem description. This isfollowed by a concept model, which is then transferredinto a computer model. Frequent validation requires aconstant adaptation of the concept and computer model[17]. In the context of this paper, a first version of thecomputer model will be presented besides the conceptof the digital twin, so that insights into the followingresearch questions can be gained:

• How should a suitable system architecture bedesigned?

• Which data are required for the digital twin?

• At what frequency should the parameters of anorder picking system be reviewed?

• Which decisions should be made automatically bythe digital twin in the future?

3.4. Systems architecture

The digital twin was intended to be used for differentorder picking systems and to be easily adaptable tomodifications in the system. For this reason, aservice-oriented architecture was chosen. The digital

twin for order picking systems currently consists ofseven services (shown in figure 2).

All data is stored in the TwinDa service. Thestorage is object-oriented. For example, one genericdata object is generated for each picking area in thesystem - this contains information on layout, orders,resources and the stored assortment. When customizingthe digital twin, the data from the connected data sources(e.g. Warehouse Management System) must be assignedto the object attributes once. Afterwards, the data isimported and checked automatically.

Based on the data basis, the TwinGe service createsforecasts for the order load in the order pickingsystem. For this purpose, the already known orders aresupplemented by forecast orders (time series for orders),so that the order load for the picking system can bedescribed as precisely as possible.

The parameters used in the simulation, e.g. thepicking time, must fit the values from the real systemand be adjusted in the event of deviation. TheTwinOp service regularly analyses key figures fromthe real system and compares them with the simulationparameters.

The TwinEx service monitors employeedeployment, e.g. in the order picking system andorder release. The current specifications are regularlychecked for this purpose. The TwinEx service examinesalternative decisions, taking into account the permitteddecisions. Possible decisions are examined using thesimulation and the forecast data. The TwinAn servicecontrols the simulation and evaluates the simulationresults. The simulation results are summarized inkey figures, which are used by the TwinEx service toimprove the decisions. The TwinSi simulation service isimplemented in PlantSimulation from Siemens and runson a separate workstation. The communication betweenthe services is based on a Representation State Transfer(REST) architecture. The REST standard was firstcharacterized and developed by Roy Thomas Fieldingin his dissertation in 2000 [18]. Today, most cloudproviders such as Azure, Amazon Web Services andVMware use the REST standard [19]. The use of RESTwill enable later implementation as a cloud service.A REST API consists of a client-server architecturebased on the HTTP(s)-protocol. The current versionof PlantSimulation cannot process HTTP requests.Therefore, the PlantSimulation C interface was used toaccess its own dynamic link library (DLL). The DLLprovides appropriate functions for communication withthe other services. The dataset is exchanged betweenthe services in the body of the request. The JSONformat is used to define the data.

Page 1658

Figure 2. System architecture and exemplary interfaces to external data sources

3.5. Data analysis and order forecast

The most important input value for the Digital Twinis the order data. This is permanently transferredfrom the WMS or web shop to the digital twin. Theorders contain information on the required articles(order position) and the quantity per article. Deliveryinstructions, such as the type of shipment (e.g. pallet,box) and the delivery date, are also included. Thechallenge when analyzing and forecasting order datafor a simulation run is that it is not only the orderstructure (number of positions, quantity, etc.) thathas to be determined. The time of order entry andthe requirements regarding delivery dates must also beforecast. Fig. 3 shows the procedure for analyzing andforecasting the order data.

In the first step, the order data is divided into groupsusing a density-based cluster process (DBScan). Theclustering includes the number of positions, the positionquantity, the shipping type and the delivery specification(time to delivery). Afterwards, a separate regressionanalysis is carried out for the orders of every cluster,so that the relationship between the number of positionsand the quantity per position can be described for everyorder cluster. The delivery type and the shipmenttime are determined by weighted probabilities whengenerating the order.

To determine the number of orders per day and thedistribution of incoming orders over the day, a day typeis defined first. Each day is assigned to a day type,whereby the day of the week, the month and aspectssuch as vacation times can be taken into account.

A further analysis is initiated for the combinationof a day type and an order cluster. The relevant dataset (day type/order cluster) is analyzed with regard tothe incoming orders over the day. For this analysis, theday is divided into time slots. For example, for the timeslot 10 to 11 a.m., a calculation of the number of dayson which 100 to 200 orders had already been receivedby this time period. For the relevant days (100-200orders having been received), the further inbound ordersover the day are analyzed. The result is a discretedistribution function that describes the receipt of furtherorders in the next time slot. Starting from a known initialorder backlog, different trends of incoming orders canbe described and rated with a probability of occurrence.

When generating orders for the simulation, thecurrent order intake from the connected system isqueried. The received orders are assigned to a cluster(based on parameters of the cluster center). This enablesthe order backlog to be determined for each cluster.Different order trends are calculated for each clusterand the probability of occurrence is determined. TheTwinAn service specifies which trends are used for thesimulation based on a defined service level. Usuallydifferent trends are used to investigate the robustnessof the decision. The combinations of trends for eachorder cluster result in a large number of possible loadscenarios. In order to be able to investigate as manyscenarios as possible, the order generation and thesimulation runs must perform very fast.

A very large number of historical data is requiredfor the described application. The clustering of ordersand the classification according to day types reduces the

Page 1659

Figure 3. Schematic process of trend analysis

number of data records for the individual trend analyses.The described procedure is currently being tested andvalidated. It has already been shown that the size ofthe value ranges for the trend analysis has a very largeinfluence on the accuracy.

A trend analysis without order clustering (all ordersin one cluster) describes the real order trend over theday inaccurately and does not meet the requirements.However, the simulation model which is described in thefollowing chapter relies on order data that is as accurateas possible so that the simulation results form a validbasis for controlling the order picking system.

3.6. Simulation

The simulation forms the core of the digitaltwin. Due to the high complexity of the system,the effectiveness of individual measures can only beadequately checked by a process simulation. Therelevant parameters have already been presented inchapter 3.1. In this chapter, it became clear that previousstudies, which have dealt with the investigation of orderpicking systems, do not provide information about howfrequent a parameter has to be investigated. However,since this is the core element of the simulation, Table

Table 2. Relevant parameters and their temporal

classification≤ hourly per shift daily

Human resource dimension X

Human resource distribution X

Zoning X

Batching X

Routing X

Job assignment X

2 shows the relevant parameters and a timeline for theanalysis.

Individual order picking systems can only becompared to a limited extent. This circumstancerequires a simple modelling of the real system. Thesimulation service developed for the digital twin inthe PlantSimulation software (Siemens) has an objectlibrary that can be used to quickly model different orderpicking systems. Each simulation model (order pickingsystem) has a framework that enables communicationwith the other services. Essential here are the interfacesto the TwinGe service which are used to transferthe system load (picking orders) to the simulation.Furthermore, the interface to the TwinOp serviceenables an automated check of the process parameters(e.g. the picking time in each order picking area) and theinterface to the TwinAn service allows the adjustmentof decision parameters, such as the staff assignment.The simulation results are also transferred to the servicementioned above via this interface.

Before the TwinSi simulation service can be used, itmust be parameterised in a customising process. Severalblocks are available in the object library for this purpose.The objects represent different picking methods, suchas conventional picking (person to goods) or pickingstations in combination with an automated warehouse(goods to person). The selected modules are placedand connected in the simulation model according tothe intended material flow between the areas. Figure 4shows an example system with four areas.

Whereby area 1 and 3 are passed through by pickingorders one after the other. In picking area 2 the ordersare processed parallel to the orders in areas 1/3. Finally,the picking orders are then merged in area 4. Eachindividual area (module) can then be parameterizedaccording to the real conditions. The layout parametersare set in this customizing process (for example, numberof aisles, compartments per shelf). The essential processcharacteristics are also set in the customizing process.This is necessary so that the process times of thepickers are mapped accurately. At this point we would

Page 1660

Figure 4. Structure of a multi-level order picking system

like to point out once again the close link with theTwinOp service. In the usage phase of the Digital Twin,process times should not simply be specified by the user,but should be determined cyclically from data sourcessuch as the Warehouse Management System (WMS).This is to ensure that the Digital Twin always reallyrepresents the real system. To complete the customizing,the assortment must be assigned to the order pickingareas (modules). This can also be done automaticallyvia the TwinOp service. Customizing is completedwith a functional test of the simulation model and theinterfaces.

During the development of objects in the simulationlibrary, attention has been paid to a generic design ofthe individual areas so that the modules can be usedfor as many different applications as possible. Anotheressential requirement for the development of the objectswas standardized interfaces. This refers on the onehand to the exchange of information between the objectsin the simulation model and on the other hand to thecommunication with the services already mentioned. Bystandardizing the interfaces, additional objects (pickingtypes) can be implemented if required, without havingto modify the framework or other objects.

Due to the automatic parameterization of thesimulation model with the use of the TwinOp service,the model does not have to be adapted manually by theuser after the initial customizing. Instead, the digitaltwin automatically adapts to the real state of the order

picking system, so that in this case we can speak of a”real” Digital Twin. The user only has to intervene inthe simulation model if the structure of the order pickingsystem changes. This is the case, for example, if theprocessing of orders in areas 1 and 3 is no longer serial(see fig. 4) or if another picking area is integrated intothe simulation model.

3.7. Decision support

The Digital Twin for order picking systems isdesigned to support decisions, especially in operationalissues, such as size of the picking tour, human resourceplanning and adjustments to the order-release strategy.For this purpose, the allowed decision space is definedduring the customizing of the Digital Twin. In the caseof staff resource planning, these are, for example, theminimum and maximum number of employees in anarea as well as the maximum number of employeesavailable for a certain type of process in the entireorder picking system. This defines the decision spaceof the Digital Twin. Today, order picking systemsusually have various picking areas, therefore the staffassignment must be defined for each area. But the staffassignment in one area of the picking system cannot beconsidered independently of the planning for the otherareas. One reason for this is the maximum numberof resources available for the entire system. Anotheraspect is that the number of allocated resources will

Page 1661

affect the throughput of the picking area. When ordersare processed serially in several areas, the output of onezone is the input of another zone. Due to the necessaryholistic view of the employee-deployment planningfor order picking systems, a high number of possiblecombinations for the employee assignment quicklyarise. To keep the number of necessary simulationexperiments low, a heuristic approach was developedfor the TwinEx service. This gradually evaluates certainconstellations for staff planning and uses the simulationresults to determine which further alternative shouldbe investigated. Boundary conditions are also takeninto account, e.g. that an employee must work for atleast 2 hours in one order picking area before beingassigned to another area. For the defined experimentsthe simulation runs are performed and the simulationprovides the following key figures to support decisionmaking

• throughput for each area

• lead time for each area

• order backlog for each area (incoming)

• utilization of the pickers per area

• performance per order picker per area

• service level (order in time) for each area

• lead time per order (total)

The decision regarding staff resource planning mustmeet two main objectives. It must be guaranteed that allorders have been processed on time for a given shipmentand, on the other hand, the number of pickers involvedand the necessary working hours must be as small aspossible . The TwinEx service looks for a setting forstaff resource planning that ensures a service level whichis parameterizable (mandatory criterion) and at the sametime has the lowest number of worker hours. It alsoattempts to keep the required number of employees aslow as possible over time. The different key figuresare transformed into one key figure with the help ofa cost function. The staff resource planning with thelowest costs is suggested to the user or transferred tothe control system (e.g. WMS). Different load scenarios(see chapter 3.5) can also be included in this costfunction. The parameters for the load scenarios areweighted by the forecast probability of the load scenariooccurring.

4. Implementation and evaluation

The digital twin for picking systems is being testedby a logistics service provider. The provider handles

the orders for various online shops. Order picking iscarried out in three different picking zones. In two zonesthe picking from pallet racks is realized. However, themajority of orders (73 percent) are processed in the thirdzone. These are goods to person systems. The transportof the goods is carried out by small transport robots.Figure 5 shows the warehouse of the logistics serviceprovider.

Figure 5. Warehouse of logistics service provider

Between 10,000 and 20,000 order positions areprocessed per day by 50 employees in two shifts. Thisimplementation focuses on the question of employeescheduling (Zone Allocation). This implementationfocuses on the question of human resource planning(number of workers per zone). In the context of goodsto person picking, there are two further questions: Atwhich station (port) is the picking for which onlineshop placed and how many orders should be preparedby the automated warehouse in advance. The logisticsservice provider uses an in-house developed WarehouseManagement System (WMS) with an SQL server asdatabase. The REST API actually updates the order dataevery 120 minutes. The order data are supplementedby the WMS with warehouse-specific data (e.g. storagelocations) in advanced. During the night, further dataare synchronized. These are mainly data that are usedto determine the current process times. In addition,the stock mirror is analyzed. From this informationthe current stock filling level is determined and it ischecked if all storage locations are also represented inthe simulation model (number of aisles, length of aisles).After each update the order forecast is started. For thespecified load scenarios, various options for operationare examined using the simulation service (TwinSi).The options for action were defined by the logisticsservice provider. Currently the results are written intothe database of the WMS and can be viewed by the userin the WMS. (Web interface is still under development).The user confirms or overwrites the suggestion foraction. Parameters such as the assignment of ports to

Page 1662

online stores (goods to person zone) are automaticallywritten in parameter tables in the MWS. Employees arepersonally informed about the change of a zone. Thesupervisors usually agree with the system’s suggestions.A comprehensive validation is planned in the comingmonths. Currently, the simulation run times for theperson-to-goods system are between 4.5 minutes and6.25 minutes per experiment (option to action) with 5replications. Per load scenario six predefined optionsfor action are investigated (number of open ports andallocation to online shops). The reason for the longsimulation run times is the very precise representation ofthe robot behavior in the simulation. Here the simulationmodel shall be simplified in a next step, where theeffects on the simulation results (accuracy) have to beinvestigated. The simulation times for the other areasare less than 1 minute. In figure 6 the simulation modulefor the good to person system (AutoStore) is shown.

Figure 6. Simulation modell of good to person zone

After the nightly data transfer, the TwinOp serviceuses the new data to determine parameters such as thepicking time depending on the online shop (order type).Furthermore, structural changes can be detected via thewarehouse mirror or order data. The new parametersare displayed to the user. The simulation model isupdated if the user agrees to the new parameter. Thusthe simulation model can be adjusted daily to changingconditions. The testing of the digital twin is still inthe beginning. An intensive evaluation of the orderforecast and simulation results is planned for the comingmonths. Furthermore the usability of functions shall beincreased.

5. Conclusion

The term digital twin is used in many ways. Sincea common definition has not yet been established,the published concepts are usually not ”real” DigitalTwins. Based on the definition of Kritzinger et. al.

([5]), a Digital Twin has an automated bi-directionaldata exchange. The paper clarified that Digital Twinscannot be considered separately from Big Data andData Analytics, since a large amount of data has to beprepared for Digital Twins. This processing representsone of the greatest challenges within the implementationof a Digital Twin. However, logistics and the relatedfields of application offer a high potential. The DigitalTwin concept has great potential, especially within orderpicking systems. Order picking systems are usuallyvery complex, such that the consequences of humandecisions are often impossible to estimate. In addition,individual decision criteria are subject to interactions.Furthermore, the demands on a picking system areconstantly changing, which makes it necessary toregularly review decisions.

The testing of the architecture has shown that thedata prediction for the system load is a great challenge,and that it significantly influences the results of thesimulation. With regard to the simulation, the level ofdetail a system has to be modeled to remains an openquestion. For decision support, uniform events and KPIsmust be defined that trigger a decision. This raises thequestion of the extent to which the quality of decisionstaken can be. The decision-making process wasimplemented by a cost function that takes into accountvarious key figures and load scenarios. In the next step,this methodology will be improved. Feedback loops willbe implemented, whereby the users have to evaluate aproposed decision. This is intended to replace the fixedparameterization of the cost function. Discussions withpotential users have shown that the topic of acceptanceof automated decisions is a challenge. In this respect itwill be important to prepare the traceability of a decisionmaking-process in a comprehensible way.

References

[1] M. Grieves, “Digital twin: Manufacturing excellencethrough virtual factory replication,” 2014.

[2] Q. Qi and F. Tao, “Digital twin and big data towardssmart manufacturing and industry 4.0: 360 degreecomparison,” vol. 6, pp. 3585–3593, 2018.

[3] F. Tao, J. Cheng, Q. Qi, M. Zhang, H. Zhang, and F. Sui,“Digital twin-driven product design, manufacturing andservice with big data,” vol. 94, no. 9-12, pp. 3563–3576,2018. PII: 233.

[4] C. Pettey, “Prepare for the impact of digital twins,” 2017.[5] W. Kritzinger, M. Karner, G. Traar, J. Henjes, and

W. Sihn, “Digital twin in manufacturing: A categoricalliterature review and classification,” vol. 51, no. 11,pp. 1016–1022, 2018. PII: S2405896318316021.

[6] K. Harper, C. Ganz, and S. Malakuti, “Digital twinarchitecture and standards,” 2019.

[7] A. Ashrafian, O.-G. Pettersen, K. N. Kuntze, J. Franke,E. Alfnes, K. F. Henriksen, and J. Spone, “Full-scale

Page 1663

discrete event simulation of an automated modularconveyor system for warehouse logistics,” in Advancesin Production Management Systems. Towards SmartProduction Management Systems (F. Ameri, K. E.Stecke, G. von Cieminski, and D. Kiritsis, eds.),vol. 567 of IFIP Advances in Information andCommunication Technology, pp. 35–42, SpringerInternational Publishing, 2019.

[8] G. Baruffaldi, R. Accorsi, and R. Manzini, “Warehousemanagement system customization and informationavailability in 3pl companies,” vol. 119, no. 2,pp. 251–273, 2019.

[9] M. Buckova, R. Skokan, M. Fusko, and R. Hodon,“Designing of logistics systems with using of computersimulation and emulation,” vol. 40, pp. 978–985, 2019.PII: S2352146519303059.

[10] J. Leng, D. Yan, Q. Liu, H. Zhang, G. Zhao,L. Wei, D. Zhang, A. Yu, and X. Chen, “Digitaltwin-driven joint optimisation of packing and storageassignment in large-scale automated high-rise warehouseproduct-service system,” pp. 1–18, 2019.

[11] W. Hofmann and F. Branding, “Implementation ofan iot- and cloud-based digital twin for real-timedecision support in port operations,” vol. 52, no. 13,pp. 2104–2109, 2019. PII: S2405896319315009.

[12] M. Braglia, R. Gabbrielli, M. Frosolini, L. Marrazzini,and L. Padellini, “Using rfid technology anddiscrete-events, agent-based simulation tools tobuild digital-twins of large warehouses,” in 2019IEEE International Conference on RFID Technologyand Applications (RFID-TA), pp. 464–469, IEEE,25.09.2019 - 27.09.2019.

[13] B. Korth, C. Schwede, and M. Zajac, “Simulation-readydigital twin for realtime management of logisticssystems,” in 2018 IEEE International Conference on BigData (Big Data), pp. 4194–4201, IEEE, 10.12.2018 -13.12.2018.

[14] W. Gunthner and M. ten Hompel, Internet der Dinge inder Intralogistik. Springer Berlin Heidelberg, 2010.

[15] R. de Koster, T. Le-Duc, and K. J. Roodbergen, “Designand control of warehouse order picking: A literaturereview,” European Journal of Operational Research,vol. 182, no. 2, pp. 481 – 501, 2007.

[16] T. van Gils, K. Ramaekers, Caris, and R. B. de Koster,“Designing efficient order picking systems by combiningplanning problems: State-of-the-art classification andreview,” vol. 267, no. 1, pp. 1–15, 2018. PII:S0377221717307920.

[17] M. Rabe, S. Spiekermann, and S. Wenzel, VerifikationUnd Validierung Fr Die Simulation in Produktion UndLogistik: Vorgehensmodelle Und Techniken. VDI-Buch,Springer.

[18] R. T. Fielding, “Architectural styles and the design ofnetwork-based software architectures,” 2000.

[19] T. Drilling and S. Augsten, “Konzept, aufbau undfunktionsweise von rest,” 2017.

Page 1664