Development of an ontology-driven,component based framework for the implementation of adaptiveness in a Jellyfish-type simulation model

Abstract

Simulation modelling has an ever increasing importance for complex systems. Manufacturing and related material flow or logistic systems are typical fields of application. Latest trends such as Cyber-physical systems and Industry 4.0 give a significant boost to simulation modelling as these require a digital model of the system. Complex manufacturing and related material flow systems are subject to frequent changes and pose a Big Data problem, which raises stronger requirements regarding self-adaptiveness. Conventional simulation models are to be adapted only via user interaction. Previous research steps have concentrated on the establishment of a novel simulation model structure, the so called “Jellyfish” model which unifies layout and process-type simulation models. Visualization of both aspects simultaneously enables interacting users to better understand the systems’ operation compared to the conventional models. The current paper focuses on the adaptive capability of the new model. We have concentrated on the hardest type of adaptation, the structural adaptation. In this paper, an ontology-driven component based approach is presented and explained further through an example. Application of automated ontology-matching in simulation environment is a novel approach enabling the simulation model to adapt its structure without the necessity of manual interaction.

Keywords

Simulation logistics ontologies adaptive systems

1. Introduction

In order to set up a context for the presentation of the achieved results, the first dominating trends of advanced simulation modelling are presented.

Manufacturing and the related logistic system build up a complex system. In order to be able to optimize processes, various simulations can be used. Many researchers regard it the only way to analyse complex material flow systems [20]. However, the application of the simulations is not limited to the above fields; this paper focuses on the manufacturing and logistics fields applying supply chain modelling aspects as well. Simulation models can be generally classified into System Dynamics (SD), Discrete Event Simulation (DES) and Agent-Based Models (ABM) [11]. SD models focus on flows with different flow rate among buffers with material stock. These models have the simplest operational logic and are very frequently used for modelling long term behaviour of large systems. A simple example for a stock-and-flow diagram is presented in [24] and depicted in Fig. 1.

Requirements for more detailed modelling have enabled quick spreading of Discrete Event Simulation models. Currently these are probably the most frequently used models. A DES model is generally a network of queues and servers, in which dynamic objects (entities) are streaming. DES models are extremely popular among researchers as there are many software environments which enable easy model development and implementation of complex operational logic. An example for this can be found in [3], and the model presented in the current paper has also been developed as a DES model. Tako and Robinson give a comprehensive overview and analysis on applicability of both models in the logistics and supply chain context in their paper [29].

Fig. 1.

Stock-and-flow model of an SD simulation.

There is an interesting modelling approach which unifies the advantages of SD and DES systems. Hennies et al. [16] proposed a process model utilized within mesoscopic simulation combining discrete impulse-like flows with piecewise constant flow rates instead of modelling individual flow objects. In terms of level of detail, mesoscopic simulation models fall between DES and SD simulation models. The main advantages of this simulation approach are the lower calculation efforts compared to DES simulations, and at the same time, a higher level of detail in comparison with SD simulations.

Fig. 2.

Agent Based Supply Chain Modelling and Analysis Framework [14].

Agent Based Simulation is a relatively new choice in simulation modelling, despite having existed since the early 1990s. Besides the event-controlled element of DES systems, it uses a number of autonomous agents which use predefined rules for the implementation of functionality. ABS raise the models’ complexity to a new level. An explanatory model has been presented by Harper [14]. During the research, a supply chain risk management framework has been developed that combines software agents, variable resolution agent based simulation, and a risk metrics component as depicted in Fig. 2. This model presents a military supply chain, where the aircraft agents process the failures of the aircrafts’ components. Depo agents handle the storage relevant data. Base agents are used for modelling processes of the bases where the repairs are carried out. OEM agents represent suppliers’ operation. A risk management component is used to continuously evaluate the selected, risk-related KPIs.

Fig. 3.

Role and impact of a simulation in decision making and implementation with an agent-based framework [6].

Fig. 4.

Example structure of a Cyber Physical System.

Comparing ABS models with SD and DES, it can be concluded that the simpler operational logic of an SD model is indeed a simplified ABS model, which means SD models are a subset of ABS models. DES models differ from ABS fundamentally in the activity of the model elements. ABS models apply “active” elements, as these dispose of continuously functioning behaviour. Elements of DES models are “passive” in a sense that their functionality is only activated as an event occurs (e.g. an entity exits the DES element).

Further Agent Based models can be mixed with and can be developed in the DES environment, see for example [25], a material flow application in the construction industry.

Discussing the ABS, Siebers et al. [27] underline that Agent Based simulations are capable of modelling such systems, where individuals with autonomous behaviour have an emphasized role. This is true for manufacturing and dynamic supply chains as well. These applications require a modelling of processes that is dynamic and that quickly adapts to changing requirements. ABS lets us model how people actually make decisions within a supply chain and see the effects of all decision makers on the supply chain [27].

ABS has a further advantage: the Agent Based approach in system control is widespread. Monostori et al. [21] wrote in their paper that AB techniques are welcome in manufacturing because they helped to realize important properties as autonomy, responsiveness, redundancy, distribution, and openness.

Citing a supply chain example [6], the authors give a good overview of how a simulation can react to decision problems which arise at various levels. The problem is the following: if the decision is needed at strategic level, even then tactical and operational level simulations are needed (see Fig. 3, above). These two levels together already build up a complex system, consisting of many autonomous participants who make decisions as well at their level. In that case, using an agent-based structure has great advantages because the agents’ structure resembles very much the roles of the different decision makers. In the paper’s concept there are so-called planning agents in the decision system (Fig. 3, below). These planning agents control the operating system’s agents in a 1–1 correspondence. These operating system agents are responsible for performing an agent-based discrete event simulation. Furthermore, there are agents which model the supply chain’s environment: the demand (Dm) and the incoming materials from suppliers (Su).

Despite all the real advantages of agent-based systems, we do not witness fast spreading of ABS models. There can be a number of possible causes. First, DES is frequently used by both practitioners and academics, because of the flexibility of simulation platforms. Common joint projects require a software environment which is common for both parties.

Secondly, why ABS is not spreading as quickly as expected: some people may feel a need for a decision between intelligent data content (AB systems) and realistic data content (CPS).

This is however not true, as CPS (see Fig. 4) frequently apply agents to process low-level data for the high-level KPIs. For methods of the design, modelling, simulation, and integration of cyber-physical systems, see [15].

Currently we are observing rapid evolution of CPS (Cyber Physical System) – Industry 4.0 – concepts in manufacturing and related logistic processes. This concept means the transformation of physical system components and related computational capabilities into a single system [18]. At the same time, the ever growing use of sensors and networked machines continuously generate a high volume of data (Big Data). Requirements and complexity in the areas of utilization of CPS have increased dramatically [19]. In this complexity, humans are only able to overview the system, if there are high-level key performance indicators available.

This ambition can be largely supported by development of a simulation model of the physical system, which acts as a “cyber twin”. Lee et al. [18] presented a CPS for the prediction of maintenance needs of machines. For this purpose, historical utilization patterns are used to simulate possible future utilization scenarios. Predicting future system states from historical data with the application of the digital (or cyber) twin is a core function in CPS systems.

Simulation models can play various roles in CPS. Our former research [13] concentrated on the development of a CP material flow system in which main control functionality was implemented by a simulation model.

In our opinion, there is no need to choose from SD, DES or ABS techniques, as from the above references one can see that these can be combined, so the embedded intelligence into the simulation model is only subject to programming effort.

With the rise of CPS, the main challenge for simulations is how to transform the necessary data from sensors and other computational devices.

A further challenge comes from the fact that in the manufacturing and related material flow systems, processes and material flow structures are subject to frequent changes. Thus, there are increasing requirements on the simulation model, regarding adaptiveness to the continuously altering modelled system.

Adaptiveness has multiple possible implementations in simulation models. Previously we presented the following types of adaptations [5]:

Dynamical objects in the simulation (for example: the production program of the system has changed, that is e.g. alternative routing of the streaming loading unit, but using the already established material handling processes). This is the least challenging way of adaption, the static structure doesn’t change at all, and the process times are selected from existing parameter sets, no parameter fine-tuning is carried out.

Adapting operational rules and parameters in the model (these adaptions are without changing the static structure of the model, but some process parameters are altered using actual sensory information from the physical system). Examples include continuous, automatic measurement of some material handling times which are averaged and fed back to the simulation. For a further example see [28].

Structural adaption (this means automatic reorganization or changing of the model’s static structure and parameters, if necessary).

This paper focuses on the hardest task: structural adaptation. In order to cope with this complex problem, next we present its possible solution using ontologies for a previously proposed simulation model structure (see [5]).

Findings of this paper support our general research objective which aims to develop novel simulation models. The novelty include increased openness towards both the physical processes by automatic adaptation and to the human users by better visual depiction.

2. Application of ontologies in simulation models

The introduction included expectations on the adaptive models: these should be intelligent and able to cope with CPS challenges. We added another objective at the start of our research: the model structure should be comprehensible. It comes from our experiences on the cyber-physical systems, in which humans have an increasing role in the physical part. A comprehensible model structure enables the human experts to understand better the contexts of the computer model.

To match the last requirement, our attention turned to the use of ontologies. Applying ontologies is a good methodology to abstract and concentrate information. Originating from the computer and information technology, its excellent thinking pattern gets more and more applications in other areas as well.

The term “ontology” is originally a term from philosophy and indicates the discipline that deals with existence and the things that exist. In terms of informatics, “exist” means all the subjects which can be represented by data [1]. There are different definitions of ontologies; a good example for it can be found in [30]: “An ontology is a formally defined system of concepts and relations between these concepts. Ontologies contain – at least implicitly – rules.”

Ontologies are typically constructed from the following components:

concepts (material and immaterial objects) which all belong to the scope of the ontology,

attributes of the concepts to enable evaluation actual validity of each concept,

taxonomies for the categorization of concepts, and

relations between the previously defined concepts.

According to M. Gruninger and J. Lee [12] there are three application areas of ontologies:

First of all, ontologies serve as a common base for communication and knowledge sharing between experts or even machines. This means ontologies serve as a common vocabulary of different agents.

Second, ontologies are not only used as a communication platform. They can be used directly for logic inferencing and reasoning, which means process control can be implemented with ontologies as well. Further implicit knowledge from the structure of the ontology can be drawn using the rules.

Third, ontologies can be reused for building more specific new ontologies, which means a knowledge reuse functionality.

Figure 5 presents an illustration of the creation of an ontology from database- and knowledge-related information using rulesets, which is also presented by Ni et al. in [22]. Thus this figure is symbolic, it depicts exactly that this process require mainly numeric data type information from the database and the description of the concepts’ relation.

Fig. 5.

The general framework involving data, information, and knowledge.

Applying ontologies may be beneficial for several reasons [9]. The first advantage is the dual readability for humans and machines, thus reducing the knowledge translation steps between them. Ontologies further enable inferencing, which is useful for deriving implicit facts. Inferencing means constant evolution capability: it is an agile schema management in contrast to traditional relational databases. In connection with the previous point, an automatic determination of the hierarchy is also possible. A significant advantage of ontologies is the reusability. This means that complete new ontologies can be built from inherited components.

Ni et al. [22] presented in their paper an ontology based activity modelling method in smart homes. The authors stated that there are two main approaches for modelling human activities in smart homes. These are the data-driven and the knowledge-driven approaches. Data-driven models use information from diverse sensors and other information devices. Here, besides the model’s construction, the mappings from the sensors must be established as well. Data-driven models can be beneficially applied if the incoming data is statistically uncertain. This approach has problems in the initial phase, when there are not enough training data available. Such disadvantages can be remedied by ontologies which belong to the knowledge-driven approach. In this paper, we are trying to unify the data-driven and knowledge-driven approaches by the definition of a component repository which can be continuously updated by actual data.

Fig. 6.

Classification of matching techniques [8].

As we want to change our ontology upon the current situation, there are multiple important questions to answer:

What are the principles upon which we can decide if a change is necessary?

How can we decide which ontologies are the same or more similar?

And finally, what is the appropriate way to implement ontologies into a simulation environment?

Changing of an ontology occur prominently in the simulation model’s planning operational mode, during which the model’s ontology part has to be automatically generated from components. The way of generating new components must, however, comply with the fundamental objectives of the system. Therefore, it is very important to define the system’s goals which do not depend on the system’s phase. In their paper, Diamantini et al. [7] presented a goal-oriented approach in the field of ambient assisted living (AAL), where high-level goals are described in terms of subgoals and tasks that are then linked to corresponding measures and devices. The goals can be directly measured via sensors. An example for this is the goal of fire detection. There are two tasks, temperature measurement and smoke detection, which contribute to the goal. These tasks are implemented using sensors (temperature and air transparency measurement). In case of dangerous measured values, a safety alarm is issued. In the above example, there are so called softgoals as well. These cannot be measured directly by sensors, but can influence overall operation of the system. For example, if there is a face recognized in the bathroom, this hurts the relation with the softgoal of avoiding privacy violation.

This approach seems to be very much appropriate for our simulation case, where definition of goals help to find the appropriate measures and so the best fitting ontology component. In our approach, we use system relevant measurable goals, so the KPIs must be defined as well. Softgoals could be used in logistic applications: such an objective can be “keeping the material supply uniform” or “keeping the material supply reliable” which is a complex issue to quantify. The presented example of Section 4 is however too simple to apply softgoals.

In order to implement the structural adaptive behaviour, the most appropriate elementary ontology in the component repository has to be selected. This needs a methodology which should be able to measure correspondence between the ontology of the actually sensed process and the component ontologies in the repository. This is called ontology matching. This process is illustrated for our application in Fig. 12 of Section 3.

Ontology matching has a widespread collection of published methods. In their paper [23], Otero-Cardeira et al. gave a comprehensive overview of the currently applied techniques. The matching methods are primarily classified into two categories: elementary matchers and structural matchers. The first one relates to the problem of finding concepts with the same meanings in different ontologies. Structure-level matchers find the correspondences between two ontologies by analysing how the concepts are connected in the structure.

In this paper, we will present a structure-level matcher as the used concepts are unambiguous, because these come only from pre-defined sensors.

Euzenat and Shvaiko gave an extensive overview of the ontology matching techniques in their book [8] (see also Fig. 6).

As seen in the Fig. 6, there are a lot of possibilities for ontology matching. From these, our approach belongs to the graph-based techniques. The graphs have namely much common features with the implementation in simulations. This however doesn’t exclude possible application of other, mainly structure-level methods. We concluded that our application is in the structure-level syntactic area in which taxonomy and graph-based methods can be applied. An example for taxonomy-based techniques can be found in [31]. These methods, however, cannot be applied for our case as they use taxonomies for the description of concepts, and the measurement of similarities are based strongly on them.

Graph-based methodologies are, however, more suitable for us. These regard ontologies as graphs and the matching problem as a graph homomorphism. Joslyn et al. [17] presented a method to measure the amount of structural distortion carried by an alignment between two ontologies which are represented as semantic hierarchies. The approach applies order theory, and the ontologies are represented by partially ordered sets $P = (P, ⩽)$ . Relation ⩽ is a reflexive, antisymmetric, and transitive binary relation, such as subsumption (“is-a”). We remark for better understanding that the relation in our example will be (“is following event”).

Let there be two ontologies represented by partly ordered graphs $P_{1}$ and $P_{2}$ . Define an alignment relation $F \subseteq P_{1} \times P_{2}$ , which is a collection of aligned pairs $(a_{1}, a_{2})$ , where $a_{1} \in P_{1}$ and $a_{2} \in P_{2}$ . By selecting two elements from $P_{1}$ , these are $a_{1}$ and $b_{1}$ and two elements $a_{2}$ and $b_{2}$ , from which $(a_{1}, a_{2})$ and $(b_{1}, b_{2})$ are aligned pairs, similarities of the two graphs can be calculated using the upper and lower cardinality-based distances [17]. The upper cardinality based distance ( $d_{up}$ ) can be calculated, for example, using the elements $a_{1}$ and $b_{1}$ as: $\begin{matrix} (1) & d_{up} (a_{1}, b_{1}) = | ↑ a_{1} | + | ↑ b_{1} | - 2 \cdot max | ↑ c_{1} |, \end{matrix}$ where $c_{1} \in a_{1} \lor b_{1}$ , $| ↑ a_{1} |$ and $| ↑ b_{1} |$ are upset of “ $a_{1}$ ” and “ $b_{1}$ ”. The upset is the ancestor set of the concept including the concept itself. Join of the elements is defined as: $\begin{matrix} (2) & a_{1} \lor b_{1} : = Min (↑ a_{1} \cap ↑ b_{1}) . \end{matrix}$

Similarly, lower cardinality-based distance can be calculated using the elements $a_{1}$ and $b_{1}$ as follows: $\begin{matrix} (3) & d_{low} (a_{1}, b_{1}) = | ↓ a_{1} | + | ↓ b_{1} | - 2 \cdot max | ↓ c_{1} |, \end{matrix}$ where $c_{1} \in a_{1} \land b_{1}$ , $| ↓ a_{1} |$ and $| ↓ b_{1} |$ are downset of “ $a_{1}$ ” and “ $b_{1}$ ”. The downset is the descendant set of the concept including the concept itself. Meet of the elements is defined as: $\begin{matrix} (4) & a_{1} \land b_{1} : = Max (↑ a_{1} \cap ↑ b_{1}) . \end{matrix}$

In our example, we will compare neighbouring elements. Therefore, it is enough to use only one of the two metrics, and so, we selected the lower cardinality-based distance. For details of application see Section 3.

Ontologies are already used in simulations. In their paper [2], Benjamin et al. describe the role of ontologies in facilitating simulation modelling. The authors analysed application of ontologies along the whole simulation modelling process (see Fig. 7).

Fig. 7.

Steps of the simulation modelling process [2].

During establishing the purpose and scope of the simulation model, ontologies are used to harmonize the terminology. This has a particular role if there are multiple developers in the same model. In the model concepting phase, an ontological analysis helps to map the system objects into the model and implementation of the operational logic. During the design of the detailed model, ontologies can facilitate detailed analysis and constraints.

Fig. 8.

Ontology-driven component based simulation [2].

The authors of [2] have given suggestions on an ontology-driven component based simulation framework, which starts with the creation of a component ontology. The ontologies are then stored in a virtual repository of components. Using this modelling and simulation, experts can easier identify the necessary elements, thus creating a more homogenous simulation model. The process itself is depicted in Fig. 8.

We have also used this approach with the difference that the simulation model itself should manage the component repository. On the information acquired from general ontology and particularly ontology-matching literature, we have formulated a novel type of a simulation model, which is described in details in Section 3, with particular regard to the adaptive capabilities. Summarized, the proposed Jellyfish-type simulation model belongs to the component based simulations, where the selection of the components are carried out automatically, using an ontology-matching method.

3. Description of an ontology-driven component based model with ontology matching optimization

The following section describes the so-called Jellyfish simulation model’s structural adaptation feature using an ontology-matching algorithm. The Jellyfish modelling concept has been previously presented in [5].

Here, we outline only the very basics of it. To develop this model, we made an assumption using several example simulation models from the simulation community that we can distinguish between “layout type” (for a good example see [4] and Fig. 9) and “process type” simulations (for a good example see [26] and Fig. 10). These labels correspond to the modelling focus.

Fig. 9.

Layout type simulation example [4].

Fig. 10.

Process type simulation example.

The Jellyfish model is a structured mix from these two approaches, where the processes are strictly connected to locations, and they are illustrated in the same modelling environment with the same simulation components as the layout. The Jellyfish is divided into two subcomponents (see Fig. 11). The upper part in the $(x, y)$ plane is the model of the physical layout. The other subcomponent which corresponds to the tentacles of the Jellyfish, describes the processes. This means that the modelling environment’s downward pointed z coordinate axis is indeed a time coordinate. In the actual model version, the time coordinate is not realistic, actual process lengths cannot be directly calculated from the vertical position of the simulation modelling elements.

The physical layout is constructed manually by simulation modelling experts. This must include all the possible physical locations because this part of the model is not subject to structural, only parametric adaption. The process part can, however, be subject to structural adaption.

Fig. 11.

General construction of the “Jellyfish” model.

Fig. 12.

Operational modes of the Jellyfish model.

The model can be operated in two different operational modes:

The structure building mode is the phase during which the simulation model “listens” to the real processes by acquiring data from the information devices of the modelled system. Upon the received information, the processes are actuated from the repository using an ontology matching method, described later in this section. The identified processes appear in the Jellyfish model’s process structure. This mode is thus similar to an emulation functionality which is added by a continuous process analysing function supplied by the ontologies. The operation mode’s process flow is depicted in Fig. 12(a).

The planning mode acts as a forecasting software. In this case, the manufacturing need (orders) of the future period is given. The actual state of the system from the structure building operational mode is known. In planning mode, the model decouples from the real system and tries to build up new processes. Selection among alternative processes is carried out by an agent which evaluates objective functions at the decision points. This operation mode’s process flow is depicted in Fig. 12(b).

We remark that the basic operational function of the system is the structure building mode, from which the system switches to planning mode. Switching between the two modes are carried out upon user’s direct instructions.

In structure building operational mode, matching the incoming information from the interface with the ontologies in the repository is carried out using the methodology written in Section 2.

Let there be $P_{acq}$ , an ontology of the acquired data from the information interface of a certain process. The system has already a repository of ontologies $P_{1}, \dots, P_{n}$ . As the components of the ontologies are unambiguous, because these are coming only from a well-defined interface, alignment can be carried out using a top-down approach based on the sequence of $P_{acq}$ . The matching pair from $P_{i}$ ( $i = 1, \dots, n$ ) is the element from the not yet selected elements with the same process ID and the same information content. If there are more than one element with matching information content, we select the one with the least different upper cardinality-based distance.

This way, a set of alignment relations ( $F_{i} \subseteq P_{acq} \times P_{i}$ ) are generated between the acquired data and the members of the repository. Next, for each aligned pairs, with the exception of the first aligned pair, the normalized sum of the differences between the lower cardinality-based distances is calculated by: $\begin{matrix} (5) & \begin{matrix} {Diff}_{r} (P_{acq}, P_{i}) \\ = (\sum_{t = 1}^{k - 1} | d_{low} (a_{t, acq}, a_{t + 1, acq}) \\ - d_{low} (a_{t, i}, a_{t + 1, i}) |) / N_{acq}, \end{matrix} \end{matrix}$ where $d_{low} (a_{t, acq}, a_{t + 1, acq})$ is the lower cardinality-based distance between the tth and $(t + 1)$ th aligned elements by $F_{i}$ of the acquired data. The expression $d_{low} (a_{t, i}, a_{t + 1, i})$ is the distance for the other elements of the aligned pair in the ith ontology of the repository. $N_{acq}$ is the number of elements in the acquired process ontology.

After calculating the normalized differences between the acquired data and the members of the repository, an ontology with the minimal difference can be chosen. If the difference is below a certain threshold value, then the acquired data matches the minimal differing member of the repository. In that case, this member of the repository will be inserted into the process part of the simulation model. If the difference exceeds the threshold, then the acquired process ontology gets into the repository. Later, these newly acquired processes need to be identified by process experts.

Fig. 13.

Physical part of the example model with the location of the information sources.

In planning mode, the role of the depicted agent of Fig. 12 still needs to be cleared. As in planning mode, the process structure keeps changing, so finding the optimal ontological elements of the repository is a key task. This is done by a command agent of the simulation model. This acts, if necessary, to find the optimal process ontology. The agent needs, therefore, continuous overview on the system’s defined key performance indicators. The appropriate KPIs must be selected according to the global desired performance of the system. In a material flow system, minimal labour and energy consumption, optimal stock amount, or minimal throughput time of operations can be appropriate. In order to properly select the necessary processes, the agent should know each process’ effect on the changing of the KPIs’ values. Therefore, in the structure building mode, the software calculates KPI values at the start and end of the process. This way, its effect can be determined. The changing of the KPIs is available for the Agent. Upon decision situations, the Agent checks all possible following processes, regarding their effect on the KPIs. The process with the most suitable changing effect on the KPIs will be selected from the repository. Next, the planning mode continues using this selected process. Operation of the Jellyfish model will be further explained through an example.

4. Test results

In the following, a computer simulation model of a simplified manufacturing and related material handling (logistic) system is demonstrated and explained. The model has been developed using Siemens Tecnomatix Plant Simulation Version 12.2.0. The example manufacturing system consists of a single machine which can produce product A, B, or C. Each product is assembled from parts 1 and 2, however, in different quantities. The production plan consists of a continuously changing mixture of the three products which causes the parts’ demand fluctuating.

Parts 1 and 2 are transported by towing trucks to the respective buffer areas in loading units (see Fig. 13). There are three different supply processes which can be started periodically at each 60 minutes. Supply process I means transport of two loading units of part 1 to the respective buffer area. Supply process II designates transport of two loading units to the buffer area of parts 2. Supply process III is a mixed one, where the towing truck transports a single loading unit of part 1 and another of part 2 to the respective buffers. All three processes use the same route (see “original supply route” in Fig. 13.

Though in the example we examined only the effect of the route change, the model itself is capable of handling all types of adaptation which relates to events and objects having location and time information. This can be location change of machines and buffers, so the adaptation can effect technological processes as well.

Fig. 14.

Implementation of a transport process in the simulation model.

The physical system is equipped with the following devices which have information connection to the simulation:

A command agent to start the route of the towing truck. It gives a unique identifier to the transport process and indicates the type of the process (supply I, …, III).

A scheduling device, which initiates a supply process if necessary. On the signal of the scheduling device, the command agent starts a process which can be a transport process or a wait command (in the case if the stock at the buffers is currently sufficient, and therefore, no parts’ supply is needed). In our example, the scheduling device works at a 60 minute frequency.

Identification (ID) devices, which confirm information on the delivery of parts at the buffers.

Position signals, which give information on the passing transport vehicles. These are depicted in Fig. 13 by Pos1 to Pos7.

Information feedback on the completion of each assembled product.

Information content from the devices include the identifier of the device, a timestamp on the occurrence of the event, and the process ID to which it belongs. Positions of the information sources are presented in Fig. 13. The example model has only a single type of resource: the transporter. There are two of them in the system.

There are altogether three types of processes in the system:

Assembly process: It appears in the Jellyfish model at the position of the machine and represents the machine’s operation. After finishing the process, the free signal always requires another Assembly process as the next process.

Wait process: On the signal of the scheduling device, the command agent gets a signal from it. In order that it does not remain a free signal, the command agent must start a process. If it is not necessary to transport any of the parts, a Wait process starts which end after some time automatically. It is a closed process, so there remains no free signal at the end.

Transport process: If parts are needed in the buffers at the time of the scheduling device’s signal, the command agent starts the most appropriate transport process. These can be supply I, …, III processes. Supply I is a transport process of materials to Buffer 1. Supply II transport materials to Buffer 2. Supply III is a combined route supplying both buffers. Implementation of transport process I in the simulation model can be seen on the left side of Fig. 14.

Next, a closer look is taken on the details of this example process of Figs 13 and 14. The first step of the process is Process start. It gets a signal from the command agent. Its location is at the warehouse. Next, there are further steps which end by position signals Pos1 and Pos2. Afterwards, as the Transporter reaches Buffer 1, an identification process step takes place during which the material gets into the buffer from the Transporter. This process step is located at Buffer 1. Finally, there are further position process steps as the Transporter travels back to the Warehouse along the path: Pos3, 4, 5, 6, 7.

We have described before that the model can construct itself in planning mode if objective functions of the model are defined. In this example model, there are two objective functions.

Table 1

Calculation of objective functions for processes

Process	Effect on stock 1	Effect on stock 2	$O_{1}$ value before	$O_{1}$ value after	Utilization, $C_{act}$	$O_{2}$ value	Change of $O_{total}$ during the process
Supply I	$+ 2$	$+ 0$	0.7	0.5	0.6	0.4	$- 0.08$
Supply II	$+ 0$	$+ 2$	0.7	0.9	0.6	0.4	$+ 0.08$
Supply III	$+ 1$	$+ 1$	0.7	0.7	0.65	0.35	0
Wait	$+ 0$	$+ 0$	0.7	0.7	0	1	0

First, the stocks in Buffer 1 and 2 must be kept at optimal level. If the optimal stocks in the Buffers are $S_{1 opt}$ and $S_{2 opt}$ respectively, the objective function can be formulated as follows: $\begin{matrix} (6) & \begin{matrix} O_{1} = & max {(1 - \frac{abs (S_{1} - S_{1 opt})}{2 \times S_{1 opt}} \\ - \frac{abs (S_{2} - S_{2 opt})}{2 \times S_{2 opt}}), 0}, \end{matrix} \end{matrix}$ where $S_{1}$ and $S_{2}$ are actual levels of stocks (number of unit loads) in the buffers. Value of $O_{1}$ equals to 1 if both buffer levels are optimal. If one or both the stock levels differ too much, the value can decrease to zero.

Second, the energy consumption of the Transporter elements must be as low as possible. This objective function differs from the previous one because it can only be calculated with respect to a reference time interval. The objective function is so formulated that it gives a smaller value if the energy consumption in the reference time period is large. Thus, for each Transporter for a reference time, the function can be calculated as follows: $\begin{matrix} (7) & O_{2} = \frac{C_{\max} - C_{act}}{C_{\max}}, \end{matrix}$ where $C_{\max}$ is the maximal consumption of the vehicle for the reference period, which means it runs with maximal utilization. $C_{act}$ is the actual energy consumption for the reference time period.

The overall objective function can be calculated as the multiplication of $O_{1}$ and $O_{2}$ : $\begin{matrix} (8) & O_{total} = O_{1} \times O_{2} . \end{matrix}$

As mentioned before, goal functions are used for the construction of a simulation model. Let’s see an example for this. The scheduling device gives a request for a possible transport or wait process. In this case, the command agent has a choice to start any of supply I, II, III, or a wait process. The optimal process from these to select can be determined using the goal functions. To do this first for each optional process, its effect for the changing of the objective function can be calculated. In our example, let’s suppose that the optimal stock levels for the buffers are: $S_{1 opt} = 5$ pieces and $S_{2 opt} = 5$ pieces, respectively. Let’s suppose, there are actually the following stock levels in the buffers: $S_{1} = 6$ pieces and $S_{2} = 3$ pieces. As we know exactly how each transport process increases the stocks and also how much energy consumption it requires, we can calculate the total objective function values for each option. Using these values, the option with the maximal objective function should be selected, and the respective simulation elements should be created in the model.

Details of the example calculations include Table 1. Based on the calculated values, the program itself is able to decide that it creates the ontology of Supply II into the model and starts it.

The next example shows structural adaptation in the structure building mode. Let’s suppose that in the previous time interval, transporters drive along all supply routes through positions Pos1, …, Pos7.

In the case of transporting Part 1, it would be better to turn after Pos3 to Pos6 with respect to the travel time (and so to the energy consumption). The company’s experts realized it and implemented the necessary process changes (modified route in Fig. 13). Next, it will be explained how the simulation will be able to realize that a process has been changed and what the reaction will be.

Fig. 15.

Operational modes of the Jellyfish model.

In structure building mode, the information from the physical system is tried to be matched with the stored ontologies in the repository. This is depicted in Fig. 15. The acquired information from the physical system is illustrated on the right side on a blue background. The arrow symbolizes its matching with the previously acquired ontologies. The ontologies from which the program chooses is shown in Fig. 15. In the figure, the lower cardinality-based distances are depicted as well (see “Diffr” value in Fig. 15). There is a preset threshold value of 0.1. Because all the differences are larger than this, the model decides that it is a new process ontology.

Therefore, a new process has been created in the repository. Afterwards, this new process is applied during the structure building of the simulation model.

5. Summary and further research

This paper has presented a novel simulation modelling structure which is particularly suitable for adaptive modelling besides its advantageous visual features. Unification of a layout type and a process type simulation model, however, looks overcomplicated at first sight. Especially the process part with the description by the simulation elements seems to be a too complex modelling method, compared to e.g. description of logic parts which are hidden in the processing elements’ internal parametrization and coding part. Depending on the number of actually running processes, the number of simulation elements can be much more than the number of the ones belonging to the layout part. A further critical comment can be that in case of several parallel running processes, we lose the overview as the processes cross each other many times.

Summarizing the above possible problems, we can formulate that Jellyfish models may let too much information into the simulation model’s space and visualize it directly, therefore, the capability of overview can be hard.

The advantages of the presented approach exceed the possible problems. The most important is that a large amount of information is intentionally and continuously let into the simulation model. In our point of view, the processing of information and the optimization of processes directly inside the simulation environment open new possibilities. We believe that the automatic generation of process ontologies in the simulation model’s $x, y, t$ space lets us apply automatic ontology optimization algorithms. The resulting optimized ontologies will help to select such combination of processes which cause optimal operation of the physical system. In our approach, the continuous connection of space and time in the Jellyfish model will enhance the capability of the optimization method.

At the same time, human operators must avoid losing the capability to overview the process ontologies. This is intended to be prevented by the use of higher-level ontologies.

As seen from the above, we are currently in the middle of a research, but the area is so novel that even a solution to a single problem – the structural adaptation – is worth publishing. In the forthcoming research, we keep our focus on the simulation models space. Therefore, the next task of the research will be focusing possibilities for optimization of the generated ontologies inside the simulation model. Thus, according to our expectations, optimal process chains can be suggested for the manufacturing and logistics systems automatically.

References

Ay, Context modeling and reasoning using ontologies, University of Technology Berlin, 2007.

Benjamin,

Patki and

Mayer, Using ontologies for simulation modeling, in: Proceedings of the 2006 Winter Simulation Conference, 2006.

Bohács,

Gáspár,

Kádár and

Pfeiffer, Simulation support in construction uncertainty management: A production modelling approach, Periodica Polytechnica Transportation Engineering44(2) (2016), 115–122. doi:10.3311/PPtr.8547.

Bohács and

Kulcsár, Reconstructuring information of damaged barcodes using various prediction models, in: Proceedings of the 10th International DAAAM Symposium “Intelligent Manufacturing & Automation: Past–Present–Future”, Vienna, Austria, 1999, pp. 53–54. ISBN 3-901509-10-0.

Bohács and

Rinkács, Development of a novel material flow simulation model for the integration of spatial and process relevant information, Logistics Journal: Not Reviewed (2016). doi:10.2195/lj_NotRev_bohacs_en_201604_01.

L.A.

de Santa-Eulalia,

J.-M.

Frayret and

D’Amours, Essay on conceptual modeling, analysis and illustration of agent-based simulations for distributed supply chain planning, INFOR: Information Systems and Operational Research46(2) (2008), 97–116. doi:10.3138/infor.46.2.97.

Diamantini,

Freddi,

Longhi,

Potena and

Storti, A goal-oriented, ontology-based methodology to support the design of AAL environments, Expert Systems with Applications64 (2016), 117–131. doi:10.1016/j.eswa.2016.07.032.

Euzenat and

Shvaiko, Ontology Matching, Springer, Berlin, 2013. doi:10.1007/978-3-642-38721-0.

Feilmayr and

Wöß, An analysis of ontologies and their success factors for application to business, Data & Knowledge Engineering101 (2016), 1–23. doi:10.1016/j.datak.2015.11.003.

10.

Franconi, Revised ontology task handbook, Deliverable TONES-D07, 2007.

11.

J.P.

Garcia-Sabater,

Lloret,

J.A.

Marin-Garcia and

Puig-Bernabeu, Coordinating a cooperative automotive manufacturing network – An agent-based model, in: Cooperative Design, Visualization, and Engineering, 2010, pp. 231–238. doi:10.1007/978-3-642-16066-0_34.

12.

Gruninger and

Lee, Introduction, Communications of the ACM45(2) (2002), 39–41. Special Issue: Ontology applications and design. doi:10.1145/503124.503146.

13.

Gyimesi and

Bohács, Developing a new logistics based model and pilot system for construction, Periodica Polytechnica Transportation Engineering43(4) (2015), 211–217. doi:10.3311/PPtr.7845.

14.

T.J.

Harper, Agent based modeling and simulation framework for supply chain risk management, Ph.D. dissertation, Air Force Institute of Technology, Wright Patterson AFB, OH, USA, 2012.

15.

Hehenberger,

Vogel-Heuser,

Bradley,

Eynard,

Tomiyama and

Achiche, Design, modelling, simulation and integration of cyber physical systems: Methods and applications, Computers in Industry82 (2016), 273–289. doi:10.1016/j.compind.2016.05.006.

16.

Hennies,

Reggelin,

Tolujew and

P.-A.

Piccut, Mesoscopic supply chain simulation, Journal of Computational Science5(3) (2014), 463–470. doi:10.1016/j.jocs.2013.08.004.

17.

C.A.

Joslyn,

Paulson and

White, Measuring the structural preservation of semantic hierarchy alignments, in: Proceedings of the 4th International Conference on Ontology Matching (OM’09),

Shvaiko,

Euzenat,

Giunchiglia,

Stuckenschmidt and

Noy, eds, CEUR Workshop Proceedings, Vol. 551, CEUR-WS.org, Aachen, Germany, 2009, pp. 61–72.

18.

Lee,

Bagheri and

H.-A.

Kao, A cyber-physical systems architecture for Industry 4.0-based manufacturing systems, Manufacturing Letters3 (2015), 18–23. doi:10.1016/j.mfglet.2014.12.001.

19.

Leitão,

A.W.

Colombo and

Karnouskos, Industrial automation based on cyber-physical systems technologies: Prototype implementations and challenges, Computers in Industry81 (2016), 11–25. doi:10.1016/j.compind.2015.08.004.

20.

Malindžák,

Straka,

Helo and

Takala, The methodology for the logistics system simulation model design, Metalurgija49(4) (2010), 348–352.

21.

Monostori,

Váncza and

S.R.T.

Kumara, Agent-based systems for manufacturing, CIRP Annals – Manufacturing Technology55(2) (2006), 697–720. doi:10.1016/j.cirp.2006.10.004.

22.

Ni,

Pau de la Cruz and

A.B.

García Hernando, A foundational ontology-based model for human activity representation in smart homes, Journal of Ambient Intelligence and Smart Environments8(1) (2016), 47–61. doi:10.3233/AIS-150359.

23.

Otero-Cerdeira,

F.J.

Rodríguez-Martínez and

Gómez-Rodríguez, Ontology matching: A literature review, Expert Systems with Applications42(2) (2015), 949–971. doi:10.1016/j.eswa.2014.08.032.

24.

Park,

Sahleh and

S.Y.

Jung, A system dynamics model for the simulation of the management of a water supply system, in: 2nd International Conference on Geological and Civil Engineering, IPCBEE, Vol. 80, IACSIT Press, Singapore, 2015. doi:10.7763/IPCBEE.

25.

Rinkács,

Gyimesi and

Bohács, Adaptive simulation of automated guided vehicle systems using multi agent based approach for supplying materials, Applied Mechanics and Materials474 (2014), 79–84. doi:10.4028/www.scientific.net/AMM.474.79.

26.

Rostkowska, Simulation of production lines in the education of engineers: How to choose the right software?, Management and Production Engineering Review5(4) (2014), 56–65. doi:10.2478/mper-2014-0036.

27.

P.O.

Siebers,

C.M.

Macal,

Garnett,

Buxton and

Pidd, Discrete-event simulation is dead, long live agent-based simulation!, Journal of Simulation4(3) (2010), 204–210. doi:10.1057/jos.2010.14.

28.

Song and

N.N.

Eldin, Adaptive real-time tracking and simulation of heavy construction operations for look-ahead scheduling, Automation in Construction27 (2012), 32–39. doi:10.1016/j.autcon.2012.05.007.

29.

A.A.

Tako and

Robinson, Comparing the use of discrete-event simulation and system dynamics models, in: 2007 Winter Simulation Conference, 2007.

30.

Voß, Begriffssysteme – Ein Vergleich verschiedener Arten von Begriffssystemen und Entwurf des Integrierenden Datenmodells, HU Berlin, 2004.

31.

Warin, Using WordNet and semantic similarity to disambiguate an ontology, Technical report, Institutionen för lingvistik, Stockholms Universitet, 2004.