Digital Twin of an industrial bottling line

This thesis work aims to build a virtual solution of a Digital Twin of a production line for a customer whose core business is the bottling of liquid solutions. All the equipment involved in the production line sends its system monitoring data to a central system that collects it. In the initial phase of the thesis, a data analysis and data mining study are presented that aim to characterise the company's data. Furthermore, this data exploration phase is preparatory to modelling the Digital Twin's virtual solution. Indeed, the theory behind the Digital Twin was studied during the data analysis, defining the possible applications in the analysed context. Furthermore, the thesis presents a bottling line simulator developed by applying the queuing theory to model the system. In addition, a dashboarding tool is presented to show a possible solution for the connection between the virtual and physical environment of the Digital Twin. The data analysis study allows to define and characterise of the data owned by the customer and defines the methodologies used to extract the data collected by the system already implemented in the bottling line. Moreover, the data mining process does not underline any relationship between collected data and line performances. That is due to the fact there are no standard methodologies to collect data, and line workers can manually change data. Hence, to study the system in depth, it is important to define new standards to collect data and implement new sensors on machines that are able to measure how operators set machines during changes of the bottle format to produce. Conversely, thanks to the analysis of historical data, it is possible to define the parameters describing the operation of a machine in the virtual environment during a simulation. The parameters are described by random variables representing service time, the time between failures and recovery time, and characterised by probability distributions that are derived based on the historical data of work periods. The simulator performance is tested over a never-before-seen work period. The mean percentage error (MPE) of the number of bottles produced varies from 16% to 24%. In addition, the thesis presents an optimisation of the simulator that allows running multiple simulations in parallel by exploiting all the machine's cores. The simulator was first tested on a machine with four cores, and the time to run 70 simulations with a single process took 32 minutes. Conversely, by exploiting all the cores, the time is halved.