Scholarlink

<p>To address the challenges of low accuracy in modulation recognition of multiple radar signals in complex environments and the difficulty in efficient retrieval, this study proposes a radar signal modulation recognition and hash retrieval method based on SimCLR contrastive learning. This method constructs a multi-scale gated residual network (MSGRNet) for feature extraction. By designing a gated residual unit (GRU), integrating a convolutional block attention mechanism (CBAM), and combining a multi-scale global fusion module (MSGF), it effectively fuses local details and global context information in the signal, enhancing feature extraction for different modulation signals. On this basis, a self-supervised learning framework is adopted, using unlabeled samples for model pre-training, and then connecting to the classification network and hash coding network, respectively, to achieve modulation recognition and hash retrieval of radar signals. Experimental results show that, across varying training sample sizes and signal-to-noise ratio settings, the proposed method performs excellently in the recognition and retrieval of 12 radar signal types, with the highest recognition accuracy reaching 98.13% and the retrieval mAP remaining above 92%, both outperforming similar supervised learning methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MSGRNet: Self-supervised contrastive learning for radar signal modulation recognition and hashing

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Li Chen,
  • Jianhua Cheng,
  • Mingming Li

摘要

To address the challenges of low accuracy in modulation recognition of multiple radar signals in complex environments and the difficulty in efficient retrieval, this study proposes a radar signal modulation recognition and hash retrieval method based on SimCLR contrastive learning. This method constructs a multi-scale gated residual network (MSGRNet) for feature extraction. By designing a gated residual unit (GRU), integrating a convolutional block attention mechanism (CBAM), and combining a multi-scale global fusion module (MSGF), it effectively fuses local details and global context information in the signal, enhancing feature extraction for different modulation signals. On this basis, a self-supervised learning framework is adopted, using unlabeled samples for model pre-training, and then connecting to the classification network and hash coding network, respectively, to achieve modulation recognition and hash retrieval of radar signals. Experimental results show that, across varying training sample sizes and signal-to-noise ratio settings, the proposed method performs excellently in the recognition and retrieval of 12 radar signal types, with the highest recognition accuracy reaching 98.13% and the retrieval mAP remaining above 92%, both outperforming similar supervised learning methods.