*3.3. Use of Metamodelling in Laboratory ZIMS*

In the laboratory, ZIMS is used for the determination of individual holon action metamodelling. An essential part of using the simulation as support for control is to train the simulation network with

data. For the verification of metamodeling, a model of manufacturing cell of the ZIMS concept is created, which consists of three machines. They gradually work on intermediate A, which enters the system at regular intervals, every 3 minutes. It is then transported by a conveyor belt to the buffer of the first machine, S1, from which it is taken and subsequently worked on if the machine is free. When the operation is completed, the semi-finished goods are transported again using the conveyor belt into the buffer of the next machine, S2, and the procedure is repeated. The transport between the last machine, S3, and the "output" is not considered.

The operational times of the machinery and the times of transport between the workplaces (Table 1) are the same at a purely theoretical level; in the event of a failure to run the system, the machines work at 100 (which is unrealistic, but it is only an explanation of the process of working in the formation of the metamodel). However, overall system productivity is affected by the failure of the second machine, S2, which occurs at particular time intervals *X1* = {*x11,x12,...,x17*} = {*15,20,25,30,40,50,60*} and the repair time is defined by a set of *X2* = {*x21,x22,...,x26*} = {*5,8,10,12,15,20*}.

**Table 1.** Times, transport times and the interval of arrivals of intermediate products into the system.


The denotation of variables is X1—time between failures; X2—repair time; Y—lead time of production.

Then we selected (based on short pilot runs) time simulation, namely, one working week with a single-shift 7.5-hour operation (i.e., 2250 min) and a time of production of 50 min. After completing all these steps, we could proceed to the implementation of simulation experiments.

These input data were performed for all combinations of the levels of factors X1 and X2 mentioned above, which totals 42 simulation runs (Table 2).


**Table 2.** Results of simulation experiments.

Subsequently, after verifying the data from the simulation, we determined the ones that will serve to train the network and those that will be test data. Artificial neural networks (ANNs) were also tested during the learning process, and validation was not necessary [33]. For training, we selected a set of 35 combinations of data obtained from simulation runs Table 3. The training set Table 4 modified the scales, and a generating error was detected using the test set. The entire process of creation, training, testing and validation took place in the Matlab program environment [34].


**Table 3.** Training data (inputs and outputs) for the artificial neural network (ANN).

**Table 4.** Test data for the ANN.


After we enter all the input factors and commands to display an error between the outputs of the ANN and the specified Y results, rendering the ANN output differences for the test data and the actual output of YT, a network training order with the training data is entered. When starting the learning process, a Figure 13 window appears, which can be followed by the training process, the number of running eras, the duration of learning, and a shrinking/increasing error [35].


**Figure 13.** Example of the ANN training process.

Once the ANN has been trained and reaches a satisfactory result, it is possible to use the ANN to address specific problems by putting new values into ANN for which no simulation runs have been made, but the responses to them are of interest to us. The validation itself does not need to be carried out since test data have already been used. However, for a demonstration, in Table 5, we compare outputs generated by our ANN for specified input data with simulation outputs.


**Table 5.** Comparison of the results of the simulation and trained ANN.

#### **4. Discussion**

Based on the knowledge learned from the long-term research in the field and the practical experience gained in dealing with the projects in the industry, we can anticipate the development of simulation environment requirements for the factories of the future.

Due to the fast onset of solutions of Industry 4.0 and the extensive use of sensors the main task for future simulation environments is the ability to model and simulate the behaviour of complex systems. When using a large number of sensors, processing data in real-time and the autonomous behaviour of the elements of the manufacturing system, the factories of the future will experience emergent phenomena.

This change will cause the simulation systems today to be used to simulate an emerging complexity. Therefore, one of the crucial tasks for the creators of simulation systems will be to develop solutions to simulate complexity in manufacturing systems. For the dynamism of the autonomous behaviour of the elements of manufacturing systems, the principles of multi-agent systems can be used in simulations, which today represents the agent simulation. Another suggested development will be the effort to "simplify" complex problems, in which way, one of the routes can be the use of simulation metamodelling. Several types of research work addressed in our department declare this development trend.

One of the crucial requirements for a new simulation environment will be its ability to offer the functionality of the emulatory environment. The integration of the real manufacturing system with its digital and virtual models will enable both offline and online optimisation, and the simulation will become part of real-time control systems.

The future simulation environment will naturally reflect the requirements of the factory of the future. In its creation, all modelling and statistical support tools, which are now commonly used in the simulation, will be used.

However, this will fundamentally change the way the simulation is implemented. Three main approaches to simulations (event orientations, process orientation and activity orientation) have traditionally been used, while new simulation algorithms will be built on distributed, autonomous principles. Due to the requirement for the reconfigurability of manufacturing systems, new simulation systems will have to offer entirely new functionalities and thematic templates, as shown in the example of the research and development of the agent simulator for future hospitals or the development of a multi-agent control (and simulation) system of complex logistics systems.

Methods and tools supporting the transformation of physical systems into virtual ones are evolving. The dynamics of the development of such systems are displayed in Figure 14.

**Figure 14.** Evolution of real system to virtual model [36].

As to further perspectives for the development of modelling and simulation in factories of the future, the Department of Industrial Engineering sees, in the following period, further integration of the various industrial engineering methods and tools used in the industry into computer simulation software tools. The concept of planning and control of future factories, through the use of computer simulations, forecasting of demand and monitoring of enterprise performance indicators (e.g., productivity), must also be developed within the framework of the approach mentioned above. These themes enter into a new dimension because the coordination of the production chain is becoming a prevalent task for all stakeholders, and the aim is to achieve a common synergy effect.

The research conducted has clearly shown that the theory of complex systems, on the basis of the requirements of factories of the future, has progressed considerably and is already providing practical tools for designers of future manufacturing systems. Key contributions of the research base on case applications are:


In the frame of research limitation, it can be said that we mainly focused on an application on the reconfigurable manufacturing system and the processes within (e.g., manufacturing, logistics). However, the model can be used on conventional manufacturing systems that have a certain level of communication ability and self-awareness.

We focused our current efforts on researching new approaches to simulating complex systems using agent simulation. In the future, we want to focus our research work mainly on the area of intelligent manufacturing systems, which using reconfigurable manufacturing systems, adaptive logistics and the concept of competency islands.
