Introducing ML to IMA Technology - System Perspective

Machine learning (ML) technologies hold substantial promise for revolutionizing the aerospace industry, offering advancements in health system monitoring, autonomous systems, and drone-based operations. However, their integration into aerospace systems introduces significant safety challenges, necessitating stringent regulatory guidance to ensure technology reliability and safety. Depending on input data complexity, high-performance hardware like Graphical Processing Unit (GPU)s, TPUs, and Application-Specific Integrated Circuit (ASIC)s are crucial for ML applications, yet aviation's stringent safety standards restrict their usage. This paper focuses on the critical aspects of system architecture, particularly in the context of leveraging the well-established Integrated Modular Avionics (IMA) for ML execution. Challenges in effectively integrating ML within IMA systems are discussed, along with various ML integration strategies, including embedding ML in applications, hosting models on the same hardware, or utilizing external hardware for specialized tasks. Case studies are explored to understand practical implementations and future research directions to enhance ML integration within the IMA framework, emphasizing the evolving role of TinyML in enabling efficient, low-power ML implementations in aerospace applications.