Scholarlink

<p>To address the challenges of large model size and insufficient feature representation in automatic modulation recognition under complex electromagnetic environments, in this paper, a deep learning model based on multi-channel feature fusion is proposed. and a lightweight self-attention mechanism. A multi-branch input network is constructed to extract the I/Q sequence and two additional single-sequence features of the signal. A lightweight self-attention module is designed, incorporating a channel grouping mechanism and a sparse connection strategy to efficiently utilize spatial and channel features while reducing computational complexity. On the public dataset RML2016.10a, the proposed model reduces the number of parameters by approximately 60% compared to the MCLDNN model and achieves a maximum classification accuracy of 93.2%. Classification performance is significantly improved while a lightweight architecture is maintained by this approach, offering an effective solution for real-time signal processing in requiring lightweight implementations.</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

A lightweight multi-channel feature fusion approach for automatic modulation recognition

  • 期刊论文

  • 发表日期：

    2025年9月19日

查看全文
  • Shaohang Gu,
  • Jia Su,
  • Chunjing Liu,
  • Xiaoyu Cheng

摘要

To address the challenges of large model size and insufficient feature representation in automatic modulation recognition under complex electromagnetic environments, in this paper, a deep learning model based on multi-channel feature fusion is proposed. and a lightweight self-attention mechanism. A multi-branch input network is constructed to extract the I/Q sequence and two additional single-sequence features of the signal. A lightweight self-attention module is designed, incorporating a channel grouping mechanism and a sparse connection strategy to efficiently utilize spatial and channel features while reducing computational complexity. On the public dataset RML2016.10a, the proposed model reduces the number of parameters by approximately 60% compared to the MCLDNN model and achieves a maximum classification accuracy of 93.2%. Classification performance is significantly improved while a lightweight architecture is maintained by this approach, offering an effective solution for real-time signal processing in requiring lightweight implementations.