Design and implementation of embedded subsystems for monitoring and managing vital parameters

This dissertation presents an interdisciplinary research pathway that integrates functional safety engineering, biosensing technologies, and eXplainable Artificial Intelligence (XAI) for embedded and resource-constrained systems. The common objective across these research lines is the development of intelligent and autonomous systems that are not only functional and efficient, but also interpretable, verifiable, explainable and safe. In the current landscape of pervasive computing and embedded intelligence, ensuring reliability and transparency has become essential, particularly for monitoring vital parameters of -- natural or artificial -- complex systems in domains such as transportation, healthcare, and human–machine interaction, where incorrect or opaque behavior can have severe consequences. The first research line investigates functional safety within railway automation systems, focusing on the validation of autonomous train operation (ATO) over the European Train Control System (ETCS). As automation in railway control advances, it becomes increasingly necessary to guarantee system integrity, redundancy, and predictability under all operating conditions. Through systematic hazard analysis and its implementation in a Vital Control (VC) module, the study shows how formal safety requirements can be met on embedded platforms, ensuring deterministic behavior and compliance with IEC 61508 principles. Indeed, the module is demonstrated to ensure deterministic safety responses, such as autonomous emergency braking and system decay management, when remote supervision fails or subsystem faults occur. The implemented architecture supports a hierarchical safety logic that preserves both functional continuity and fail-safe behavior. The experimental validation confirms that formal safety engineering principles can be effectively realized on embedded platforms, achieving real-time operation with limited computational resources. The second research line investigates biosensing technologies aimed at continuous and non-invasive physiological monitoring. In particular, the study explores chloride ion detection as a proxy for sweat analysis, due to its relevance in thermoregulation, hydration assessment, and early detection of clinical conditions. A complete wearable biosensing platform is designed and implemented, integrating the electrochemical sensor, conditioning electronics, microcontroller-based acquisition unit, and embedded signal processing chain. The system’s design emphasizes modularity, low power consumption, and robustness against environmental variations. Experimental results demonstrate the system’s capability to perform stable and repeatable measurements of chloride concentration. Furthermore, the platform provides a solid basis for the future integration of intelligent data analysis modules, enabling context-aware health monitoring and adaptive physiological feedback. This contribution positions wearable biosensing not only as a biomedical tool, but as a component of a broader human-centered cyber-physical ecosystem, where sensors, intelligence, and explainability are interconnected. The third and central contribution concerns the work performed on MAFALDA 2.0, a framework for XAI on MCU. Unlike most state-of-the-art ML systems that rely on opaque models and cloud-based computation, MAFALDA 2.0 enables on-device learning and reasoning while preserving transparency and interpretability. Implemented entirely in C language, the framework supports Incremental Learning and Federated Learning paradigms, allowing models to evolve dynamically over time and across distributed devices. Its core innovation lies in the integration of symbolic reasoning and semantic matchmaking, which transforms the classification process into a reasoning task grounded in ontology-based knowledge representations. This approach enables both global and local explainability, as each decision can be expressed in terms of symbolic relationships between input features and learned concepts. Moreover, MAFALDA 2.0 includes a configurable cutoff parameter that regulates the depth and granularity of the user-readable explanations. Extensive experimental evaluation, conducted on standard datasets for monitoring vital parameters, shows that the proposed framework achieves accuracy comparable to state-of-the-art TinyML models while maintaining a small memory footprint, short inference times, and low energy consumption. The system supports on-device training and prediction without relying on external computational infrastructures, thus enabling fully autonomous intelligent behavior even on resource-constrained hardware. Overall, the dissertation advances the state of the art toward trustworthy and self-explanatory embedded intelligence, where functional safety, biosensing, and explainability converge to enable reliable human–machine symbiosis in safety-critical, wearable and ambient intelligence contexts.