Learning-Capable Aviation Safety Systems for Real-Time Anomaly Detection

Flight Data Recorders (FDRs) continuously capture operational flight data, yet their information is predominantly used for retrospective accident investigation. The objective of this study is to present and evaluate a learning-capable aviation safety monitoring framework that learns nominal landing-trajectory behavior from successful flights and identifies persistent deviations from this learned behavior during an ongoing flight. The proposed intelligent flight data recorder (I-FDR) framework uses historical flight data to predict future trajectory evolution and compares predicted behavior with learned successful landing boundaries. The study considers the Waiting Period as the interval between the identification of a trajectory anomaly and subsequent violation of a predefined safety constraint such as minimum descent altitude. The framework is evaluated using four tailored deep learning architectures on real-world ADS-B trajectory data from two independent historical controlled flight into terrain accident cases. Results demonstrate that the framework can identify deviations from nominal behavior before safety constraints are violated, while predictive performance varies with prediction horizon and model configuration. These findings highlight the importance of evaluating the temporal margin available after anomaly identification, in addition to predictive and computational performance, for online aviation safety monitoring.