MATF-Net: Multiscale Attention With Tristream Fusion Network for Radar Modulation Recognition in S-Band

Abstract

Automatic modulation recognition of radar signals plays a crucial role in information-centric warfare and holds significant importance in military applications such as radar detection. In modern information-centric battlefields, the S-band (2–4 GHz) is widely utilized for tasks such as pulse radar detection due to its abundant spectral resources, excellent adaptability, and suitability for equipment miniaturization. However, under electromagnetic countermeasure conditions, radar signals within the S-band become dense and complex, making accurate modulation recognition particularly challenging. Existing methods often fail to adequately extract and fuse the multimodal features of signals, resulting in unreliable recognition performance under complex electromagnetic environments. Consequently, achieving robust AMR of radar signals under low signal-to-noise ratio (SNR) conditions has become critically important. To address the aforementioned challenges, this article proposes a multiscale attention with tristream fusion network to improve automatic modulation recognition of radar signals under low SNR conditions, aiming to mitigate issues such as feature ambiguity and noise interference. The proposed network comprises three main components: the three-stream feature extraction network (TFEN), the self-attention fusion network (SAFN), and the multiscale information fusion network (MIFN). Within TFEN, three specialized modules are designed—the spatial extraction module, the corresponding extraction module, and the temporal-compensation module. By parallelly extracting spatial features, amplitude and phase information, and temporal compensation features, TFEN effectively addresses the performance degradation issues typically encountered in low SNR scenarios. The SAFN and MIFN modules prioritize salient information across different modalities, compute interfeature correlations, and perform weighted fusion to enable dynamic selection and multiscale integration. This enhances the representational capacity of the fused features. Simulation results demonstrate that the proposed model achieves an average accuracy of 88.54% across SNR levels ranging from −20 dB to 20 dB, significantly outperforming existing methods and exhibiting superior adaptability.