Towards the Monitoring of Operational Design Domains using Temporal Scene Analysis in the realm of Artificial Intelligence in Aviation

Abstract

The application of Artificial Intelligence (AI) in aviation has gained significant attention in recent years, particularly in safety-critical domains such as aviation. One possible application in this domain is the next version of the Airborne Collision Avoidance System X (ACAS X). The current system, ACAS II, with its state-of-the-art implementation, TCAS II, has been in use for decades and has significantly reduced the number of mid-air collisions. However, it relies on a simple rule/heuristics-based logic, leading to false positives which increases the workload of both pilots and air traffic controllers unnecessarily. The next generation of collision avoidance systems, ACAS X, instead uses exhaustive lookup tables to determine resolution advisories. These lookup tables, however, are too large to be stored on current avionics hardware. Thus, neural networks can compress these lookup tables significantly, enabling the deployment of ACAS X on current avionics hardware. Nevertheless, deploying an AI-based system for predicting resolution advisories raises safety concerns regardless of whether it is used in commercial or unmanned aircraft. To mitigate these concerns, it is crucial to not only train the AI-based system on a substantial amount of data but also to understand and define the environmental conditions in which is supposed to operate. This concept is referred to as the Operational Design Domain (ODD). Monitoring ODD conditions is essential to ensure the safe operation of the AI-based system. This paper presents a novel methodology for predictive ODD monitoring, leveraging temporal scene analysis to assess potential scenarios that an AI-based system may encounter. Temporal scene analysis is a methodology that analyses scenarios by dividing them into discrete scenes, each representing a specific point in time within the scenario. This approach allows for a detailed examination of the scenario’s progression, identifying critical situations and transitions that may impact the AI-based system’s performance and safety. For this, it utilizes a database of scenarios generated based on the ODD description. Splitting the scenario into scenes and rearranging them to create new synthetic scenarios increases the data available for predictive ODD monitoring. All this will be demonstrated in the context of a collision avoidance use case.