AI Safety Verification for Aviation Collision Avoidance Systems

Integrating AI into safety-critical aviation systems, such as collision avoidance, demands rigorous verification to meet stringent safety standards set by bodies like EASA. A key challenge is ensuring that the data used to train and verify these AI systems accurately represents all possible operational scenarios. This paper introduces a structured engineering process to define target data distributions, model parameters, and quantitatively assess the representativeness of AI/ML operational design domains. The proposed method guides developers through identifying suitable target distributions and evaluating coverage results against EASA's learning assurance objectives. It specifically examines statistical measures, finding Kullback-Leibler divergence and Cramér's V suitable for large aviation datasets, unlike the chi-squared test. Demonstrated with AI-based airborne collision avoidance using experimental data, the approach shows how statistical comparisons can bolster safety-by-design principles for AI. This contributes to a systematic AI engineering process that aligns with evolving aviation safety guidelines.