Robust and reliable Fault Detection and Isolation (FDI) is crucial for the certified, critical embedded systems onboard civil aircraft. For such aerospace flying vehicles, traditional industrial practices rely heavily on a combination of data-driven and model-based approaches, providing dedicated monitors to achieve straightforward fault isolation.

In this context, this presentation will show that transitioning toward learning-driven designs, and especially Artificial Intelligence (AI) techniques, would introduce significant technical and regulatory challenges. Due to the scarcity of the faults to be detected, AI supervised machine learning is currently constrained by severely unbalanced data classes, though it remains viable for developing offline-tuned, deterministic virtual sensors. Unsupervised techniques enable novelty detection but impose heavy real-time computational burden during dimensionality reduction and anomaly detection, while requiring a secondary fault identification step Reinforcement learning remains incompatible with current certification and real-time constraints.

This talk will also illustrate that from an industrial perspective, deploying AI-driven FDI requires resolving core issues regarding data representativeness and the limitations of massive data collection on older aircraft fleets. Compliance with the EASA AI roadmap for systems is mandatory, necessitating robust technical guarantees on explainability, determinism, and i.a. uncertainty quantification. Ultimately, future implementations must strictly limit computational burden while delivering a highly efficient detection rate with an extremely low false alarm rate to avoid unnecessary, capability-degrading aircraft reconfigurations.

Throughout the presentation, concrete industrial examples will be provided to illustrate these technical challenges.