Evaluation of YOLO for Automatic Radar Waveform Detection and Classification in Congested EME

Modern Radar Electronic Support Measures (RESM) systems need to deal with increasingly complex signals embedded within challenging Radio Frequency (RF) environments. This paper proposes a method of generating Pulse Descriptor Words (PDWs) based on the “YOLO” neural network image processing algorithm. The YOLO method is applied on spectrograms and estimates the modulation scheme observed and predicts a bounding box to localise the signal, allowing the pulse-width and bandwidth of the signal to be estimated. The method is demonstrated on spectrograms generated from both simulated and experimental datasets. The ability to detect and classify pulses in the presence of co-channel interference was demonstrated using real intercepts collected in a congested electromagnetic environment (EME) with a 0.41 % probability of false alarm. Good detection and classification performance was also experimentally proven down to -9 dB SNR with around 8% absolute percentage error in bandwidth and pulse width estimation with.