On the Exploration of Temporal Fusion Transformers for Anomaly Detection with Multivariate Aviation Time-Series Data

In this work, we explored the feasibility of using a transformer-based time-series forecasting architecture, known as the Temporal Fusion Transformer (TFT), for anomaly detection using threaded track data from the MITRE Corporation's Transportation Data Platform (TDP) and digital flight data. The TFT architecture has the flexibility to include both time-varying multivariate data and categorical data from multimodal data sources and conduct single-output or multi-output predictions. For anomaly detection, rather than training a TFT model to predict the outcomes of specific aviation safety events, we train a TFT model to learn nominal behavior. Any significant deviation of the TFT model's future horizon forecast for the output flight parameters of interest from the observed time-series data is considered an anomaly when conducting evaluations. For proof-of-concept demonstrations, we used an unstable approach (UA) as the anomaly event. This type of anomaly detection approach with nominal behavior learning can be used to develop flight analytics to identify emerging safety hazards in historical flight data and has the potential to be used as an on-board early warning system to assist pilots during flight.