Detecting Controllers' Actions in Past Mode S Data by Autoencoder-Based Anomaly Detection

Abstract : The preparation and execution of training simulations for Air Traffic Control (ATC) and pilots requires a significant commitment of operational experts. Such a mobilisation could be alleviated by a decision support tool trained to generate a realistic environment based on historical data. Prior to studying methods able to learn from a dataset of traffic patterns and ATC orders observed in the past, we focus here on the constitution of such a database from a history of trajectories: the difficulty lies in the fact that past flown trajectories are properly regulated, that observed situations may depend on a wide range of potentially unknown factors and that ownership rules apply on parts of the data. We present here a method to analyse flight trajectories, detect unusual flight behaviours and infer ATC actions. When an anomaly is detected, we place the trajectory in context, then assess whether such anomaly could correspond to an ATC action. The trajectory outlier detection method is based on autoencoder Machine Learning models. It determines trajectory outliers and quantifies a level of abnormality, therefore giving hints about the nature of the detected situations. Results obtained on three different scenarios, with Mode S flight data collected over one year, show that this method is well suited to efficiently detect anomalous situations, ranging from classic air traffic controllers orders to more significant deviations. Detecting such situations is not only a necessary milestone for the generation of ATC orders in a realistic environment; this methodology could also be useful in safety studies for anomaly detection and estimation of probabilities of rare events; and in complexity and performance analyses for detecting actions in neighbouring sectors or estimating ATC workload.
Complete list of metadatas

Cited literature [20 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-02338690
Contributor : Véronique Soullier <>
Submitted on : Wednesday, October 30, 2019 - 9:48:20 AM
Last modification on : Saturday, November 16, 2019 - 11:31:39 PM

File

DTIS19175.1571068581_preprint....
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02338690, version 1

Collections

Citation

Xavier Olive, Jeremy Grignard, Thomas Dubot, Julie Saint-Lot. Detecting Controllers' Actions in Past Mode S Data by Autoencoder-Based Anomaly Detection. SESAR Innovation Days 2018, Dec 2018, SALZBURG, Austria. ⟨hal-02338690⟩

Share

Metrics

Record views

33

Files downloads

13