Development and Integration of Virtual Sensors for Reliability Monitoring of High-Performance Hybrid Propulsion Systems

The new 2026 technical regulations significantly change the architecture and management of the Power Unit. The greater contribution of the electrical part, the absence of the electric turbocharger and the interaction between the internal combustion engine and the more powerful electric motor compared to the previous years, determine new monitoring, reliability and control needs. However, the usage of physical sensors is limited by mass, cost, and installation constraints introduced by the new regulation. Virtual sensors could represent a possible solution to measure quantities not directly measurable, reconstruct missing signals and implement forms of analytical redundancy for diagnostic purposes. This thesis proposes to develop and validate virtual sensors for an F1 Power Unit with particular attention to engine reliability monitoring applications and the integration with an already developed tool for diagnostic. This thesis tries also to answer a question: To what extent can a virtual sensor reconstruct relevant Power Unit parameters and contribute to reliability monitoring under different operating conditions? To answer this question different AI models with different capabilities to replicate non linearities and temporal dependencies are compared. The validation methodologies do not consider just metrics but also model performance during transients, robustness across different operating regimes and the ability to generalize to conditions not included in training. The comparison between models is also conducted separately for each physical quantity with the aim of identifying a relationship between the dynamics of the phenomenon, the characteristics of the sensor, and the most suitable machine-learning architecture. The results demonstrate that the most suitable architecture strongly depends on the physical nature of the target quantity. Variables dominated by quasi-static relationships can be accurately reconstructed using simpler architectures such as MLPs, whereas quantities affected by thermal inertia, transport delays, and long-term memory benefit from temporal models such as LSTMs and TCNs. The study also confirms the importance of physics-based input selection, which often provides more robust and generalizable results than purely automated feature selection procedures. The developed virtual sensors were subsequently integrated into an existing anomaly detection system to generate operating-point-dependent references and dynamic diagnostic thresholds. Experimental results show that the selective integration of five virtual sensors maintains diagnostic performance close to the baseline configuration. The independent integration of seven complementary virtual sensors reduces false positives but also introduces false negatives, whereas the simultaneous integration of all twelve virtual sensors produces a more substantial degradation in classification performance. These findings demonstrate that virtual sensors can effectively support Power Unit reliability monitoring, provided that their design, validation, and integration are performed consistently with the underlying physical phenomena and the requirements of the complete diagnostic chain.