Offline Radar Pulse Track Association with Deinterleaving Errors

Sammanfattning

Radar systems play a crucial role in detecting the position and velocity of objects by emitting electromagnetic signals into the environment. In electronic warfare, it is essential to identify which pulses from pulsed radar systems originate from the same system, in order to track the activity of the radar system. This problem can be solved by deinterleaving the radar pulses, a process that aims to sort the received pulses within a window into different tracks based on which radar system generated the pulses. However, typically deinterleaving does not process the entire recording at once, as such sets of sequences are generated from the deinterleaving process. Additionally, deinterleaving is prone to errors, necessitating solutions to associate sequences to form complete tracks and correct for the errors. This thesis investigates time-series clustering as a solution to this problem, proposing three distinct algorithms: Mixture of Hidden Markov Models Clustering, a Fisher-Rao distance based clustering algorithm and a Long Short-Term Memory based Autoencoder. These algorithms represent varied approaches to time-series clustering, including model-based, distance-based and representation-based methods respectively. Additionally, a two-step process is proposed for error correction: identifying errors within each track using Out-of-Distribution detection, followed by inserting incorrect pulses into the correct track with a classifier. Among the algorithms studied, Mixture of Hidden Markov Models Clustering demonstrates the best performance for sequence association, albeit with high processing costs. However, Fisher-Rao and the Long Short-Term Memory based Autoencoder also yield promising results, offering potential for extension to online applications. For error correction, a combination of a Isolation Forest followed by a Decision Tree was observed to perform well under these circumstances.