Ensuring the Trustworthy Development of AI-Based Applications Compatible to Future EASA Regulations

Kurzfassung

Due to the increased capabilities of artificial intelligence (AI) in recent years to solve a wide variety of tasks, the aviation industry sees the introduction of AI into systems as a viable and important solution for current and upcoming challenges. However, due to the paradigm change to a data-driven development, the previous certification framework relying on a requirement-driven development is unsuitable for certifying AI-based systems. To solve this issue, the European Union Aviation Safety Agency is currently developing guidelines for the development of certifiable AI-based systems. In its guidelines, EASA describes different objectives that developers must achieve to provide evidence that the AI system is trustworthy and can be certified. One central part of this trustworthiness argument is the learning assurance framework. The learning assurance is EASAs equivalent to the development assurance for traditional software development. Central to the learning assurance are the definitions of a Concept of Operations (ConOps), an Operational Domain (OD), and an AI/ML constituent Operational Design Domain (ODD). The latter is defined to describe the operating conditions for the AI/ML constituent and to be used for the data management. While EASA specifies certain properties concerning these concepts, no guidance for developers on the format or methodology to define these concepts is given. Therefore, this thesis introduces a methodology to define these concepts, satisfying the requirements outlined by EASA. Concerning the concepts of the OD of the overall system and the AI/ML constituent ODD, a tabular definition language is introduced based on ISO 34503:2023. Furthermore, processes were introduced to define the different necessary artifacts. For the specification process of the AI/ML constituent ODD, different preexisting steps were identified to incorporate the data-specific considerations, that must be included in the AI/ML constituent ODD. These steps include the projection of system-level attributes to AI/ML constituent attributes based on the perception of the system and the identification of domain-specific concepts for the AI/ML constituent. To validate the methodology, it was applied to the pyCASX system, an open-source implementation of an airborne collision avoidance system, where the goal is to introduce machine learning to reduce the memory footprint. Using the proposed methodology, a ConOps and an OD for the pyCASX system were defined, and individual AI/ML constituent ODDs for the two subsystems, VCAS and HCAS. The methodology proved it was able to produce an AI/ML constituent ODD of finer detail compared to other ODDs defined for the same airborne collision avoidance use case. Furthermore, it was shown that most of the requirements for the considered objectives of EASA can be achieved. Thus, the proposed novel framework is an important step toward a holistic framework following EASA's guidelines contributing significantly to current research.