Radar Pulse Deinterleaving with Transformer Based Deep Metric Learning

TL;DR

This work tackles radar pulse deinterleaving when the number of emitters is unknown. It proposes a metric-learning pipeline that uses a sequence-to-sequence transformer trained with triplet loss to produce emitter-discriminative pulse embeddings, which are then clustered non-parametrically to recover emitter partitions. The method is evaluated on synthetic data using extrinsic clustering metrics such as AMI, ARI, and V-measure, with the transformer achieving strong performance (e.g., AMI ≈ 0.882) and outperforming a GRU baseline and an identity clustering approach. The study highlights practical considerations, including over-splitting at low emitter counts and the need to extend to longer sequences and real data, outlining directions for future work.

Abstract

When receiving radar pulses it is common for a recorded pulse train to contain pulses from many different emitters. The radar pulse deinterleaving problem is the task of separating out these pulses by the emitter from which they originated. Notably, the number of emitters in any particular recorded pulse train is considered unknown. In this paper, we define the problem and present metrics that can be used to measure model performance. We propose a metric learning approach to this problem using a transformer trained with the triplet loss on synthetic data. This model achieves strong results in comparison with other deep learning models with an adjusted mutual information score of 0.882.

Figures (6)

• Figure 1: The radar pulse deinterleaving problem requires partitioning a set of radar pulses by the emitter from which the pulses originated. Here the colours represent the emitters and the features vectors $x_{i}$ denote PDWs. This can be approached using metric learning where we generate an embedding for each pulse in the context of the whole pulse train and cluster these embeddings using an off the shelf clusterer.
• Figure 2: The architecture for the training (a) and inference (b) pipelines share the common component of a sequence-to-sequence transformer used as a pulse embedding model. We use vanilla dot product attention and no positional encodings. (a) During training we use the triplet loss to push apart embeddings of pulses from different emitters and bring together those from the same emitter. (b) During inference we use a clustering algorithm (HDBSCAN) to generate an integer label $l_i$ for each embedding from which we can trivially determine the predicted partition for the set of pulses.
• Figure 3: A plot of the different PDW features against time of arrival for an example pulse train in the test dataset. The pulse train contains 13 emitters with significant overlap in parameters. The colour of each data point indicates which of the emitters the pulse originated from.
• Figure 4: A confusion matrix of the true number of emitters in the synthetic test set against the number of clusters in the partition predicted by our transformer. The black line represents an ideal scenario where the predicted number of clusters is the same as the true number of clusters. It can be seen that the transformer generally predicts there to be more clusters than there are emitters.
• Figure 5: The true number of emitters against the mean AMI for that number of clusters. The shaded area represents the bootstrapped Bootstrap$0.1-0.9$ confidence interval. It is notable that the width of the confidence interval is much larger for high numbers of clusters where we have few data points (The distribution of data points can be seen in Figure \ref{['fig:confusion-matrix']}).