基于多通道轻量化的自动调制识别网络  

An automatic modulation recognition network based on multi-channel lightweighting

作　　者：万进华 尚俊娜[1] 张华娣 WAN Jinhua;SHANG Junna;ZHANG Huadi(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China;The 36th Research Institute,China Electronics Technology Group Corporation,Jiaxing 314033,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018 [2]中国电子科技集团公司第三十六研究所,浙江嘉兴314033

出　　处：《电信科学》2025年第2期41-56,共16页Telecommunications Science

基　　金：浙江省教育厅一般科研项目(No.Y202455360);浙江省“十四五”第二批研究生省级教学改革项目(No.JGCG2024202)。

摘　　要：自动调制识别技术在无线通信领域具有十分重要的作用。现有的自动调制识别模型在识别精度上表现出色,但大多数方法在参数量与模型性能之间难以实现理想的平衡。针对该问题,设计了一种多通道融合的轻量化调制识别(multi-channel lightweight modulation recognition,MCLMR)网络。MCLMR网络将幅度、相位、频率以及零中心归一化瞬时幅度的谱密度最大值作为输入。使用可分离卷积模块巧妙地组合4个输入,从而深入挖掘这4个输入的空间相关性。设计了基于时间衰落多头自注意力(multi-head self-attention,MHSA)机制结合门控循环单元(gated recurrent unit,GRU)的GRU-MHSA(gated recurrent unit-multi-head self-attention)模块进一步提取时间相关性。可分离卷积模块与GRU-MHSA模块的结合在空间维度与时间维度提取信号特征。在基准RML2016.10a数据集上的仿真结果表明,所提方法优于其他9种典型方法。在2~18 dB信噪比下平均识别精度达到92.39%,最高识别精度达到93.36%,这说明MCLMR不仅参数量少,计算复杂度低,在识别精度上也表现出色。Automatic modulation recognition technology plays an important role in the field of wireless communication.Existing automatic modulation recognition models perform well in recognition accuracy,but most methods have difficulty in achieving an ideal balance between the number of parameters and model performance.To solve this problem,a multi-channel lightweight modulation recognition(MCLMR)network was designed.The amplitude,phase,frequency,and maximum spectral density of the zero center normalized instantaneous amplitude were taken as inputs by the MCLMR network.A separable convolution module was used to cleverly combine four inputs to dig deeper into the spatial correlation of the four inputs.The gated recurrent unit-multi-head self-attention(GRU-MHSA)module multihead self-attention(MHSA)based on time fading and gated recurrent unit(GRU).to further extract the temporal correlation.Signal features in spatial dimension and time dimension were extracted by the combination of separable convolution module and GRU-MHSA module.Simulation results on the benchmark dataset RML2016.10a show that the proposed method is superior to other 9 typical methods.At 2~18 dB signal-to-noise ratio,the average recognition accuracy reaches 92.39%and the highest recognition accuracy reaches 93.36%.This shows that MCLMR not only has a small number of parameters and low computational complexity,but also has excellent performance in recognition accuracy.

关 键 词：调制识别 轻量化网络 多模块融合 多头自注意力机制 

分 类 号：TN92[电子电信—通信与信息系统]