Flight trajectory pattern generalization and abnormal flight detection with generative adversarial network

Flight trajectories expectedly differ from each other, yet they follow a set of patterns that are flown previously and optimized for different conditions. These patterns may or may not be trivial and can be influenced by a number of effects, including airspace utilization, con-trollers’ cognitive complexity, weather, NOTAMs. Due to the high variance, it is a rather compelling task to accurately classify trajectories into a desired set of categories based purely on its statistical properties. Therefore, we propose a hybrid methodology to generalize over the flight trajectories and decide whether they are abnormal or not: a) with a more statistical approach considering the time-based features of the trajectories, and b) with a more pattern based approach. The statistical approach utilizing the LSTM autoencoders enables to train the model with the history and rapidly predict the class of the flight, inherently considering the time-based features of a flight trajectory. The pattern generalization is achieved via a Generative Adversarial Network (GAN), which generates samples that look like real trajectories. The abnormality detection in the patterns is done with an autoencoder part of GAN that learns how to reconstruct typical sequentially dependent trajectory data after applying dimension reduction. While LSTM autoencoders enable to capture the class of a flight with short time windows, GAN predicts pattern anomaly with larger window observations. Hence, this heterogeneous approach allows us to sweep the airspace with asynchronous frequencies. The obtained results show that the proposed architecture is quite capable of the abnormality classification of the trajectories as it accomplishes to accurately detect simulated fighter aircraft trajectories in commercial flight heavy airspaces, and the generative model provides good utility to capture the pattern-based distribution of flight trajectories.