Development and integration of Machine Learning models for optimising production processes in Software-as-a-Service solutions

The thesis is set in the context of Industry 4.0 and the Horizon Europe AIDEAS project, addressing the use of Machine Learning techniques for the optimisation of production processes within Software-as-a-Service solutions for operational decision support. The overall objective is to investigate the potential of Artificial Intelligence methodologies in production scheduling, with a particular focus on their applicability in real and complex industrial contexts. The first part consists of a systematic review of the literature, supported by bibliometric analyses on Scopus and Web of Science, in order to identify the main AI techniques used in scheduling problems, in particular Particle Swarm Optimisation, Neural Networks and Reinforcement Learning, and the benefits reported in the industrial field. However, the analysis highlights a limited availability of documented case studies based on real industrial applications, confirming the relevance of the empirical contribution of this research. The experimental part of the thesis includes two industrial case studies developed in collaboration with European manufacturing companies. The first case concerns Multiscan, a Spanish company operating in the assembly of fruit and vegetable sorting machines. The scheduling problem is modelled as a Hybrid Flow-Shop with integrated Open-Shop block, including constraints related to both machines and operator assignment. The scheduling problem is formulated as a Markov Decision Process and solved using Deep Reinforcement Learning, allowing for integrated management of the sequence of operations and human resources. The results show that the learned policy outperforms traditional approaches, such as the FIFO rule and a metaheuristic based on Ant Colony Optimisation, in terms of average delay, operational stability and inference times, allowing the production plan to be recalculated in less than ten seconds. The second case study has been carried out at PAMA S.p.A., an Italian manufacturer of boring machines, and focuses on optimising the machining process for cast iron columns. In this context, two artificial neural networks are designed and trained to predict material removal parameters, with the aim of reducing rework while maintaining very tight dimensional tolerances. The dataset is constructed from actual process measurements and undergoes a thorough pre-processing and correlation analysis phase. The Multi-Layer Perceptron models achieve errors in the order of microns and, in 85% of validation cases, suggest lower removal than the traditional method, indicating a potential improvement in term. Finally, the thesis proposes a conceptual model of a "Provider Planner" based on Deep Reinforcement Learning within a Manufacturing as a Service architecture, capable of dynamically managing orders from multiple customers and adapting almost in real time to changes in demand and the availability of production resources.