| 摘要: |
| 宽带侦察场景下，虽然信号检测识别准确率高，但计算资源消耗过大的问题亟待解决。为此，提出了一种基于多尺度卷积注意力的稀疏检测方法（Multi-scale Convolution Attention Sparse Detection,MSCA-S。该方法结合信号时频图的先验知识，通过建模信号在时间轴上的远距离依赖关系并抑制频率轴的无关干扰，设计了多尺度水平卷积注意力机制（Multi-scale Horizontal Convolution Attention,MSHCA，联合提取信号的多维特征，有效提升检测识别精度，并通过水平卷积降低模型计算复杂度。基于MSHCA，构建了层次化堆叠的宽带信号检测方法，利用稀疏特征参数进一步减少计算资源需求。在频谱范围为2.5 MHz的青岛实采及仿真宽带信号数据集上进行实验，MSCA-S在不同信噪比下的平均检测精度达95.6%，相比频率敏感宽带信号检测方法、基于Swin-Transformer的协议信号识别方法和基于101层残差网络的信号检测方法，精度分别提升了0.05%、2.94%和6.14%，计算量分别降低了1.53×1010、1.79×1010和4.59×1010。 |
| 关键词: 宽带信号检测识别 注意力机制 多尺度卷积 稀疏算法 |
| DOI：10.20079/j.issn.1001-893x.240712001 |
|
| 基金项目:国家自然科学基金重点项目（U20B2038 |
|
| A Sparse Detection Method for Broadband SignalsUtilizing Multi-scale Convolutional Attention |
| GONG An,ZHANG Jinglei,GUO Lantu,ZHAO Xiaolei,LIU Yuchao |
| (1.Qingdao Institute of Software,College of Computer Science and Technology,China University of PetroleumEast China,Qingdao 266508,China;2.China Radio Wave Propagation Research Institute,Qingdao 266107,China) |
| Abstract: |
| In broadband reconnaissance scenarios,achieving high signal detection accuracy often entails significant computational costs.To address this,a multi-scale convolution attention sparse detection(MSCA-S method is proposed,which incorporates prior knowledge of signal spectrograms by capturing long-range temporal dependencies and suppressing irrelevant frequency-domain interference.MSCA-S introduces a multi-scale horizontal convolution attention(MSHCA mechanism that jointly extracts multi-dimensional signal features,enhancing detection accuracy while reducing computational complexity through horizontal convolution.Building on MSHCA,a hierarchically stacked broadband signal detection framework is developed,and sparse feature parameters are used to further optimize computational efficiency.MSCA-S is evaluated on a real-world and simulated broadband signal dataset(2.5 MHz spectrum collected in Qingdao,achieving an average detection accuracy of 95.6% across varying signal-to-noise ratios.Compared with the frequency-sensitive signal detector,the Swin-Transformer-based protocol recognition method,and the Res-101 detection method,MSCA-S improves accuracy by 0.05%,2.94%,and 6.14%,respectively,while reducing computational costs by 1.53×1010,1.79×1010,and 4.59×1010,respectively. |
| Key words: broadband signal detection and recognition attention mechanisms multi-scale convolution sparse algorithms |
