Failure Modes or Not Failure Modes? Integrating Machine Learning in Aerospace Safety Assessment Processes

Although not yet a reality, recent developments have sparked interest in adopting machine learning (ML) across various aviation applications. Potential applications include control, health management, collision avoidance, and single pilot operations. However, integrating ML in aviation presents significant safety challenges due to its complex behavior compared to traditional software. This paper examines the impact of ML's uncertainty on aviation safety assessment processes, focusing on defining ML failure modes for learning-enabled components. Through a literature review and semantic analysis based on current aviation safety standards like ARP 4754A and its successor ARP 4754B, we identify ML-specific factors that contribute to failures of ML-enabled systems. Using a case study on multimodal visual navigation, this paper validates proposed failure modes and their contributions. Additionally, it demonstrates the application of traditional assessment methods to ML-enabled systems.