Learned Embeddings for Radar Pulse Deinterleaving : A Supervised Contrastive Learning Approach to Radar Pulse Emitter Separation Using Time-Frequency Spectrograms

Sammanfattning

Pulse deinterleaving is the task of separating a stream of received radar pulses and assigning each pulse to its emitting source. A common approach represents each pulse using a Pulse Descriptor Word (PDW), a fixed set of scalar features such as carrier frequency, pulse width, and time of arrival. While compact and well-established, PDWs can fail to distinguish pulses whose intra-pulse waveform structures differ in ways those attributes do not capture. This thesis investigates whether CNN-learned embeddings can complement PDWs by encoding waveform-level information that traditional descriptors miss. A CNN encoder is trained on complex time--frequency spectrograms of radar pulses using a Hierarchical Supervised Contrastive Loss, and two configurations are compared: a 1-level model trained to separate individual pulse signatures, and a 3-level model that additionally organises embeddings according to coarser emitter groupings. Both models are evaluated using clustering quality metrics across a range of signal-to-noise ratio conditions, including levels unseen during training. At the pulse level, the most critical for deinterleaving, both models produce highly discriminative embeddings, with the 1-level model achieving near-perfect scores and the 3-level model trading a small reduction at this level for better organisation at coarser class levels. A targeted experiment further demonstrates that the embeddings correctly separate pulses sharing identical PDW attributes but differing in waveform structure, confirming their potential as a supplementary descriptor in pulse deinterleaving pipelines.