Résumé
Automated runway detection is a rapidly growing domain in avionics. Recent advancements in the machine learning (ML) community have enabled the development of cutting-edge neural networks (NN) for this specific task. However, given the critical nature of automated flight operations, the reliability and safety of these systems must be rigorously ensured. Runtime monitoring approaches have emerged as promising solutions to enhance these properties. Monitors are components designed to continuously supervise the behaviour and outputs of NNs during runtime. They are generally categorised into two main types: rule-based and data-driven. Rule-based monitors verify that safety properties are satisfied, making them suitable for scenarios with well-defined safety requirements. In contrast, data-driven monitors are trained on antecedents to learn patterns of erroneous or unexpected behaviour. This poster explores the main challenges and preliminary solutions for designing and evaluating such monitors. Main challenges would include (1) Which types of threats should the monitor address? (2) Which kind of error detection mechanism should be used? and (3) How should the monitor be evaluated with respect to overall system safety?