Error-Driven Design of AI-Based Systems for Airborne Applications

Kurzfassung

Certification practices for airborne software have been established for decades. However, strong interest in the implementation of artificial intelligence (AI) algorithms, especially machine learning (ML), has led to a lively discussion in the community about how current practices need to be adapted to achieve safety targets. This paper focuses on an approach that combines item-level assurance practices together with architectural mitigation techniques. The uncertainty of the system behaviour is traced to the prediction errors of the ML-components. We propose a technique to analyze prediction errors through optimized testing using evolutionary methods to identify potential failure regions. Then ensemble learning is proposed as an architectural mitigation and redundancy technique to achieve the required performance. The approach is presented with a use case where a regression model is used for calculating the optimum cruising altitude.