Digital Twin and Machine Learning solutions for the Manufacturing Environment

The advent of Industry 4.0 has brought manufacturing realities to become more flexible and prone to reconfiguration in order to adapt to unexpected events and clients needs. Smart Manufacturing wants to encourage the usage of innovative technologies to promote the digital transformation, especially exploiting the possibilities offered by cyber physical systems and virtual environments (VEs). Digital Twins (DTs) have been widely adopted to virtually reproduce the physical world and to integrate real environments with their digital counterpart. The development of a DT solution for a production line can be used for monitoring activities, to assess limitations and costs of the real counterpart and to simulate enhancements before implementing possible solutions in the real world. Also, the big amount of data that flow between from the physical assets to their virtual replica can be used to train machine learning systems. Machine Learning (ML) is widely accepted as a relevant technology for industry 4.0 but it requires large datasets for training. Moreover, most ML methods require labelling, which often has to be manually entered in case of real-world data. DTs can provide a powerful instrument for training ML systems, since a simulation can generate a huge amount of data that can be automatically labelled, thus reducing the user's effort during the training dataset preparation phase. This work investigates the creation of a Digital Twin of a real production line for assembling skateboards. The proposed use case represents a complex system, which requires both the creation of a VE and the usage of a ML system. The VE has been developed with Unity 3D as an interactive environment that can be experienced through immersive virtual reality for training activities as well as for inspection or analysis activities. The DT of the line has been enhanced with YOLO (You only look once), a state-of-the-art, real-time object detection algorithm deployed on Darknet, an open-source neural network framework written in C and CUDA. The ML system has been trained with a synthetic dataset automatically generated and labelled with Blender. The proposed system allows plant designers to evaluate the benefits of introducing a novel, extremely fast and accurate object detection system on the assembly line. Moreover, the pose detection system performance enables its usage in the VE for real-time users training. The thesis is split into 5 chapters. The first chapter presents the concept of DT, its main applications and benefits for the Manufacturing sector. The second chapter contains an overview of the project requirements, along with a description of the major technologies and tools used throughout the development of this thesis. The third chapter will focus entirely on the design process and the development of the application, providing details about how the technologies have been integrated and the techniques being used. Result of this work will be presented in the fourth section, followed by a description of potential improvements and new features that could be implemented in the last chapter.