Machine Learning Algorithms for Identification of Low Signal-to-Noise Ratio Radar Signals

Abstract

Areas such as spectrum monitoring require identification of known and unknown radar transmitters to identify known and rogue users. Often such identification needs to done under conditions where the signal-to-noise ratio is low. This thesis proposes an approach to determine the unknown radar chirp parameters of a linear frequency modulated (LFM) radar waveform, assuming that the unknown parameters come from a given set of known chirp parameters. A concatenated output of matched filters corresponding to the known set of chirp parameters is presented to four well-known machine learning architectures, namely decision tree (DT), random forest (RnF), naïve Bayes (NB) and support vector machine (SVM). Realistic radar parameters for airborne, marine and weather radars were used in the simulations. The robustness of the classifiers to parameter mismatch and truncation of the radar pulse were also studied. DT outperformed the other classifiers except for the truncated pulse case (where NB and SVM performed better). RnF did not perform acceptably.