Modeling of a machine learning-based virtual copilot for helicopters

The aviation industry has always searched for new solutions to make air transport safer and more efficient. As technology develops rapidly, different needs and challenges occur. In this sense, Artificial Intelligence (AI) has opened new horizons. Exploring present and future possible AI contributions, this thesis aims to develop an onboard AI suitable for helicopter pilot co-handling to reduce pilot workload. Drawing insights from flight manuals and previous studies, it becomes evident that helicopter pilots encounter numerous critical phases during flight. It was chosen to investigate three critical use cases: predicting Vortex Ring State (VRS), detecting engine malfunctions, and aiding pilots during autorotation scenarios. Each of these cases presents unique challenges in aviation safety and operational efficiency. The different machine learning techniques are analyzed to choose the best fit for the three use cases. VRS is a complex aerodynamic phenomenon that poses a significant risk to helicopter operations. A supervised learning algorithm is developed to anticipate and provide timely warnings to pilots, enabling proactive mitigation strategies. Engine failures and malfunctions are among the most serious emergencies faced by pilots when flying a helicopter, demanding swift and accurate diagnosis. An unsupervised learning-based assistant can identify subtle deviations in engine performance, alerting pilots to potential issues before they escalate, through anomaly detection and pattern recognition algorithms. In the event of a double engine failure, the pilot has to perform an autorotative descent. This maneuver combines the difficulties of landing a helicopter with those of having to make quick decisions in a power-off flight: this dissertation explores the application of Reinforcement Learning and Imitation Learning algorithms to design a virtual assistant capable of supporting pilots during autorotation landings. The training phase of supervised, unsupervised and imitation learning relies on a dataset, collected during test flights conducted in a certified flight simulator, recreating the peculiar environment of each specific use case.