On Data-centric Diagnosis of Aircraft Systems

This article discusses results of an investigation into requirements for and problematic aspects of implementing a decision theoretic data-centric approach to the problem of data driven fault detection and diagnosis for aircraft and their subsystem, suitable for offline application to Flight Operational Quality Assurance (FOQA) type data. We discuss fundamental considerations for this type of approach, aspects of diagnosis required for a vehicle health management system, some problematic aspects of the domain, and provide a high level survey of current anomaly detection techniques, considering their suitability for diagnosis. Our principle conclusion is that for datacentric diagnosis, going beyond fault detection and localization requires a mapping from observable symptoms to diagnoses that is not readily available. This suggests a bootstrapping approach involving clustering, outlier detection and expert identification of suspected faults, providing the basis for actual diagnosis.