Curriculum
Computer Science and Innovation for Societal Challenges, XXXVI series
Grant sponsor
Fondazione CARIPARO
Supervisor
s
Tullio Vardanega
Co-supervisor
s
Anna Spagnolli

Project: A TinyML-enabled approach to embed Machine Learning in Avionics Control Systems
Full text of the dissertation book can be downloaded from: https://www.research.unipd.it/handle/11577/3516623

Abstract:The incorporation of Machine Learning (ML) into aviation infrastructure signifies a pivotal advancement, unlocking exceptional possibilities for enhancing operational efficacy, facilitat- ing predictive maintenance, and improving instantaneous decision-making. This dissertation delves into the transformative potential of TinyML, a niche of ML optimized for high-efficiency in environments with limited resources, and its capacity to overhaul control systems in avia- tion, spanning both the commercial and military spheres. It scrutinizes the complexities and methodologies for embedding ML models within the strict confines of aviation frameworks, fo- cusing significantly on addressing Out-of-Distribution (OOD) instances, which significantly jeopardize the safety and reliability of systems. A cornerstone of this research is the formulation of an innovative integration strategy for ML systems, devised expressly for scenarios characterized by limited computational resources and a high demand for energy conservation. The study pioneers a cutting-edge multi-layer early exit strategy within Deep Neural Networks (DNNs), aimed at the swift and accurate recog- nition of OOD data in real-time scenarios. By integrating a detector of minimal complexity following each DNN layer, the proposed model facilitates an immediate cessation of inference upon the detection of OOD inputs, thereby elevating the system’s computational efficiency and durability within stringent operational constraints. Furthermore, this thesis conducts an exhaustive examination of sophisticated software test- ing methodologies and strategies applicable to ML-driven systems, underscoring the critical need to customize these technologies to align with the specialized requirements of safety-critical aviation environments. Achieving a harmonious balance between model dimensions, accuracy, and functionality, the research contributes significantly to the field by delivering solutions that are both practical and reliable for these systems’ operational breadth. The outcomes under- score the superiority of the proposed model over existing alternatives, particularly regarding computational efficiency and the detection of OOD data, as gauged by AUROC. Additionally, the research expands on the critical role of Operational Design Domain (ODD) in conjunction with OOD. ODD defines the specific conditions under which a system is de- signed to operate, including environmental, geographical, and temporal constraints. Under- standing the ODD is crucial for accurately identifying OOD instances—data or situations that fall outside the system’s designed operational parameters. By closely aligning OOD detection mechanisms with the defined ODD, the research highlights how ML models can more effec- tively anticipate and mitigate risks, ensuring higher safety and reliability in aviation systems. This nuanced exploration sets a foundation for subsequent research and the evolution of ML technologies within safety-critical settings, steering towards smarter, more secure, and efficient aviation operations .