TA-LSTM: A Time and Attribute Aware LSTM for Deep Flight Track Clustering

Abstract

The rapid development of civil aviation imposes an urgent need for efficient air traffic management. Flight track clustering is the premise and foundation of air traffic control. Flight track points compose time-series data with non-identical point- to-point intervals, and the effective latent representation of track data is key to the flight trajectory clustering. Long Short-Term Memory (LSTM) is an important and popular method to extract the features of trajectory time-series data. However, LSTM ignores the irregular time intervals between track points and fails to effectively distinguish the importance of different attributes at each track point during feature extraction. In addition, the existing methods based on LSTM either ignore the distinct impacts of different attributes at each track point or simply assume that the influences of the previous track points on the subsequent ones decrease with time. In this paper, we propose a flight trajectory feature extraction unit called TA-LSTM (Time and Attribute Aware Long Short- Term Memory) and a flight track clustering model based on TA- LSTM. The flight track clustering model consists of a feature extraction layer and a clustering layer. The feature extraction layer adopts TA-LSTM to obtain the latent representation of flight tracks, which adds a time control gate and an attribute control gate to the standard LSTM unit. The time control gate manipulates the action of each track point on subsequent points to obtain the characteristics of irregular time intervals. The attribute control gat enables LSTM to capture the impact of different attributes on the flight track features at various points. The clustering layer takes the latent representation of flight tracks as the input to obtain the clustering results. The clustering model updates the parameters of TA-LSTM and the cluster centers by minimizing the Kullback-Leibler (K-L) divergence. We conduct experiments using Automatic Dependent Surveillance-Broadcast track data provided by VariFlight. In comparison with popular and state-of-the-art methods, the proposed method obtains superior pe formance in three widely used evaluation metrics, i.e., Silhouette Coefficient, Calinski-Harabaz index and Davies-Bouldin index.