Application of Machine Learning Techniques and Data-Driven Approaches to the Modeling, Control, and Diagnosis of Thermally-Driven Systems with Integrated Electric Heating

Thermal systems based on electric heating elements are ubiquitous and critically important components across a vast spectrum of application sectors. The reference market is currently undergoing a paradigmatic shift, characterized by a transition from the mere commercialization of hardware components to the provision of integrated systems and comprehensive solutions that include advanced control strategies and high-value-added auxiliary services. In this new context, the heating element ceases to be a discrete entity and is instead configured as an interdependent subsystem within a holistic architecture, whose overall performance is a function of the synergistic and dynamic interaction among its constituents. This research work, developed within the framework of an industrial doctorate, aims to address the aforementioned challenges through the formulation and experimental validation of methodologies founded on data-driven approaches and machine learning paradigms. The primary objective is the development of a digital twin—or a high-fidelity virtual representation—of the system's critical components, aimed at the accurate prediction of functional behavior in heterogeneous and non-stationary operating contexts. This approach seeks to overcome the intrinsic limitations of analytical or lumped-parameter models, which are often inadequate for accurately describing the highly non-linear dynamics and stochasticities that characterize real-world systems. A complex thermo-hydraulic system, specifically a professional coffee machine, has been selected as a test bench for the validation of the proposed methodology. This system constitutes a significant case study due to the intrinsic variability of its boundary conditions (e.g., heterogeneity of pumps and brewing units) and the non-stationarity of its process parameters (e.g., the need to generate optimal thermal profiles specific to different types of coffee capsules). Mastering such complexity is an essential prerequisite for achieving high-quality standards and ensuring process repeatability. The specific objectives of the research are twofold: (i) the synthesis of robust predictive models of the thermal system's behavior; and (ii) the use of these models for the design and implementation of advanced control algorithms capable of real-time self-adaptation to specific operating conditions. Concurrently, the research investigates the application of ML techniques for the development of ancillary diagnostic and prognostic functionalities (PHM - Prognostics and Health Management), including anomaly detection, virtual sensing of quantities that are not directly measurable, and predictive maintenance. The primary scientific and industrial contribution of this work, therefore, lies in the definition of a systematic methodological framework that enables the valorization of operational data, transforming them into a strategic asset. This facilitates the transition from a product-centric to a service-centric business model (servitization), while simultaneously ensuring the optimization of functional performance and the enhancement of the system's reliability and useful life.