Assurance-driven Design of Machine Learning-based Functionality in an Aviation Systems Context

We present an approach to the co-development of an aeronautical function integrating Machine Learning (ML) and its associated preliminary safety assurance case. Using an Autonomous Visual Landing (AVL) application as a running example, in which a Deep Neural Network (DNN) supports navigation state estimation, first we describe how we identify the contributions of ML-based functionality to system hazards. We then give the safety and ML requirements necessary for hazard mitigation, together with a candidate safety architecture aimed at mitigating the ML-induced functional insufficiencies. Thereafter, we present a structured argument embodying the rationale that provides confidence that the ML-based function will be safe for use. Collectively, these artifacts represent the core elements of a multi-viewpoint safety assurance case. We discuss how that concept is compatible with prevailing aerospace recommended practices for safety assessment, its utility to align design-time assurance considerations with operational safety needs, and how it can communicate the evidence necessary for safety assurance in an integrated way.