Radar Signal Deinterleaving With Progressive Correlation-Integrated Neural Network

Radar signal deinterleaving is a critical task in radar signal reconnaissance. Existing methods face challenges in effectively incorporating high-dimensional parameters such as the correlation parameter into end-to-end neural networks. This paper proposes a method termed “progressive correlation-integrated deinterleaving” (PCID), which is capable of processing a 133-dimensional input comprising both conventional pulse descriptor word (PDW) parameters and high-dimensional correlation parameters. PCID first constructs a correlation matrix by measuring the similarity between each pulse and its temporal neighbors, thereby introducing correlation information that captures coherent processing interval (CPI) patterns into the network. To address the high dimensionality of the correlation matrix, we decompose it into short-range and long-range components, and progressively integrate them into different stages of a convolutional baseline, aligned with its expanding receptive fields. In addition, instead of direct concatenation or addition, we introduce correlation-guided feature transformation blocks, which adaptively modulate PDW-based features using correlation-based context. Experimental results demonstrate that PCID outperforms both traditional approaches and other deep learning-based methods in deinterleaving performance, and simultaneously maintains competitive computational efficiency.