Optimizing Industry 5.0 Machine Learning-Based Applications via Synthetic Data Generation

Machine Learning (ML) innovations are revolutionizing industrial processes by improving productivity and competitiveness, particularly by creating predictive maintenance applications and Artificial Intelligence (AI) powered services. In the Industry 5.0 paradigm, which strives to reach Zero Defect and Zero Waste Manufacturing, the role of ML is crucial for optimizing these processes. However, developing ML solutions for production-ready industrial settings presents various challenges, such as data scarcity and dataset imbalance, which are further amplified when dealing with tabular data. To overcome such issues we investigate the use of various Deep Generative Models (DGMs) to generate synthetic data that closely mimics real-world conditions. Despite extensive studies of DGMs in the literature, there is still limited knowledge of their practical suitability in Industry 5.0 environments. Our research includes a detailed evaluation of several DGMs, such as Generative Adversarial Networks, Variational Autoencoders, and Diffusion Models, implemented in the gearbox assembly and testing line in the Bonfiglioli EVO plant. Based on our results, the evaluated DGMs demonstrate significant potential in generating high-quality synthetic data, that allows training a high-performance classifier to distinguish faulty gearboxes from well-functioning ones.