CLASSIFYING ADS-B TRAJECTORY SHAPES USING A DENSE FEED-FORWARD NEURAL NETWORK

Abstract

There is a recent abundance of flight trajectory data due to Automatic Dependent Surveillance-Broadcast (ADS-B) becoming a prevalent and required aviation traffic control system. Motivated by incidents like the September 11, 2001, attacks, the Department of Defense and civilian intelligence agencies have taken a renewed interest in being able to quickly flag and act on flight pattern behavior that is considered outside the norm. Due to the large volume of daily flights in the United States alone, it is almost impossible for human operators to monitor and analyze individual flights for anomalous behavior. The Department of Defense and civilian intelligence agencies stand to gain increased capability and capacity if given the ability to analyze and flag unusual flight trajectories in a matter of seconds. Anomalous behavior in many cases is determined by the overall shape of the flight pattern. This thesis uses calculated shape features to classify nine pre-determined categories of ADS-B flight trajectories using a Deep Sequential Neural Network. With a data set of 11,303 human-classified tracks, the network has performed with an overall accuracy of 71% and a categorical average F1 score of 0.33 on a validation set. It has also performed with 70% accuracy and a categorical average F1 score of 0.25 on a ten-fold cross validation. The proposed method shows promise in being able to select unusual shapes from straight trajectories and in some cases may be able to classify them.