Radar Signal Deinterleaving with Multi-Feature Semantics and Modular Network Design

Radar signal deinterleaving is essential for situational awareness in complex electromagnetic environments. Nevertheless, subtle inter-class differences among radar signal categories and large parameter measurement errors pose significant challenges to accurate deinterleaving. Existing methods often fail to comprehensively leverage the discriminative information embedded in multiple parameters. This paper proposes a radar signal deinterleaving method, termed “multi-feature semantics deinterleaving” (MFSD). First, MFSD introduces the concept of “semantics”, referring to deep features that encode discriminative characteristics from pulse streams, such as parameter values, modulation patterns and interparameter relationships. Based on this concept, MFSD decomposes the task into three modular stages: preprocessing, semantic extraction and semantic-based decision. This modular structure enables the design of flexible and extensible network architectures. Then, a concrete instantiation of MFSD is implemented using a one-dimensional convolutional neural network (1D CNN) that aligns with the proposed framework. The network incorporates a learnable normalization scheme to preserve parameter information, employs improved dilated convolutional blocks for robust semantic extraction, and utilizes attention mechanisms to enhance semanticbased decision-making. Finally, comprehensive experiments, including ablation studies and comparisons with traditional and deep learning baselines, demonstrate that the proposed method achieves over 30% improvement in mean intersection over union (mean IoU), and simultaneously exhibits superior robustness and generalization across diverse radar signal deinterleaving scenarios