Architectural Challenges in Developing an AI-based Collision Avoidance System

Abstract

Emerging trends in Advanced Air Mobility (AAM) are pushing the boundaries of the established design approaches and are forcing developers to find new ways to fulfill the need for more powerful, reliable and robust equipment for future software defined aircraft functions. Of particular interest in achieving this is the field of Artificial Intelligence (AI) and its subset of Machine Learning (ML) algorithms. The use of AI/ML within the aviation industry, however, poses significant challenges, particularly connected to safety, reliability and certifiability. This paper is about the OpenCAS, a collision avoidance system based on Feed-Forward Neural Networks. It reports hands-on experience and outlooks on systems engineering practice for ML model integration. The architectural design considerations are elaborated. Particular focus is laid on constraints imposed by the use of multiple networks within the system.