Resumo:
The use of simulation models as Digital Twins (DTs) has been standing out in recent years and represents a revolution in decision-making in production processes, being a key solution in the context of the so-called Industry 4.0. In this sense, we highlight increasingly faster and more efficient decisions from the mirroring of the behavior of physical systems through sensors, intelligent equipment, management systems and databases. The models used as DTs are updated periodically, in real or near real time according to physical changes, and provide guidelines or commands for decision making. On the other hand, despite the great applicability of this approach, challenges related to the validity of simulation models over time stand out, since traditional validation approaches do not consider the periodic update of the model. Ensuring the validity of DTs is essential, since it usually involves decisions of great impact for production systems. In addition, although it is a field of research with great importance for both researchers and professionals, we noted that there is still a gap in terms of methods aimed at monitoring the validity of DTs. Therefore, in order to contribute to the literature and fill this gap, the present work proposes an approach based on the periodic evaluation of simulation models used as DTs through Machine Learning and control chart. We suggest a monitoring tool based on the K-Nearest Neighbors (K-NN) classifier, combined with the p control chart, in order to periodically assess the validity of DT models. Initially, the proposed approach was tested in several theoretical cases in order to evaluate the functioning of the tool in situations where the physical environment differs significantly from the virtual one, a fact that would represent a possible case where the DT is not valid. In this case, data corresponding to the physical and digital environments were emulated considering standardized probability distributions. Furthermore, the tool was also implemented in two real objects of study, acting as a supplement to make DTs more robust and reliable. In this case, DTs already implemented and in the operational phase were adopted. The first object of study refers to a model that supports operational planning decisions in a medium-sized company of a clothing industry, whose processes are mostly manual. The second object of study refers to a DT implemented in an automated production cell that operates in near real time, allowing the evaluation of the main process parameters. The tool proved to be capable of monitoring the functioning of both DTs and identifying possible special causes that could compromise its results and, consequently, its validity. Finally, the broad applicability of the tool is highlighted, which can be used in different approaches of DT, including simulation models with different characteristics of connection, integration, and complexity. In this case, the proposed approach operates independently of the characteristics of the DTs, including models that operate in real or near real time, considering automated or manual physical systems and covers systems with different levels of complexity.