WVEmbs with its Masking: A Method For Radar Signal Sorting

TL;DR

This work tackles radar signal sorting (RSS) under high-density interleaving and non-ideal conditions where prior DL approaches struggle. It introduces Wide-Value-Embeddings (WVEmbs) to map PDWs into normalized, multi-periodic embeddings and a value-dimension masking strategy for hard sample mining, enabling a one-step end-to-end RSS pipeline. With a CNN-backed ModernTCN backbone, the method achieves high-accuracy, sample-level deinterleaving, notably reaching $88.680\%$ RSS accuracy and robust performance across three scenarios, while ablations show improvements over standard embeddings and enhanced training stability. The approach reduces reliance on manual preprocessing and clustering, offering an efficient, robust solution for dense interleaved radar environments and suggesting avenues for theoretical analysis and broader time-series applications.

Abstract

Our study proposes a novel embedding method, Wide-Value-Embeddings (WVEmbs), for processing Pulse Descriptor Words (PDWs) as normalized inputs to neural networks. This method adapts to the distribution of interleaved radar signals, ranking original signal features from trivial to useful and stabilizing the learning process. To address the imbalance in radar signal interleaving, we introduce a value dimension masking method on WVEmbs, which automatically and efficiently generates challenging samples, and constructs interleaving scenarios, thereby compelling the model to learn robust features. Experimental results demonstrate that our method is an efficient end-to-end approach, achieving high-granularity, sample-level pulse sorting for high-density interleaved radar pulse sequences in complex and non-ideal environments.

Figures (4)

• Figure 1: The main idea of this paper: 1. To address the low overlap probability of trivial features in interleaved radar pulse streams and the challenges in predicting and preprocessing their distributions (e.g., DOA features overlapped when signal sources are nearby, as shown in the lower left panel). WVEmbs processes and sorts the features of the original signal from trivial to useful in the "wide-value dimension", which with each dimension normalized to a 0-1 range, converted from the widely-scaled and sparse original signal, suitable for neural network processing (as shown above, note the visibility of the green radar pulses in the WVEmbs). The masking process helps the classifier focus on learning useful features, mitigating the risk of training crashes due to the predominance of trivial features (bottom right).
• Figure 2: Pipeline of our model used for RSS. The backbone is stacked ModernTCN blocks. Masks are applied with a certain probability $P$ only during model training. Experimental results indicate that $P$ has minimal impact on the model's performance.
• Figure 3: Confusion matrices for the three test scenes.
• Figure 4: Above: WVEmbs significantly enhances the stability of the training process. Value Dimension masking facilitates loss convergence and improves accuracy. Below: Model performance across an SNR range of -20 to 20 dB in three scenarios.