DNN-Based Joint Time-Domain Enhancement and Recognition Framework for LPI Radar Signals in Low-SNR Environments
Low probability of intercept (LPI) radar signals pose significant challenges for electronic support (ES) in Internet of Military Things (IoMT) due to low signal-to-noise ratios (SNRs) and complex modulations. Traditional noise reduction and recognition techniques struggle to maintain performance in low-SNR environments. Recent advances in deep neural networks (DNNs) have enabled progress in both signal enhancement and recognition tasks, but most existing methods remain dependent on time–frequency transform (TFT), increasing computational cost and relying on prior expert knowledge. Besides, these methods rarely provide comprehensive evaluation across multiple metrics. To address these issues, this article proposes LPI-ETDE-STAR, an end-to-end framework for time-domain LPI signal enhancement (LSE) and modulation recognition (LMR). The framework employs an end-to-end time-domain enhancement network (ETDE-Net) to enhance signals, which integrates convolution for local feature extraction and a structured state space (S4) module for capturing long-range dependencies (LRDs). Then, the STAR-MRNet utilizes STAR operations to model high-order and high-dimensional feature interactions for accurate modulation recognition. Comprehensive experiments show that ETDE-Net surpasses both traditional filtering and DNN-based methods for the LSE. Furthermore, STAR-MRNet achieves superior LMR performance based on the outputs of ETDE-Net, outperforming existing DNN approaches under low-SNR conditions.