| 摘要: |
| 针对复杂无线传播环境中频谱数据标签稀缺、信号传输过程中信噪比（Signal-to-Noise Ratio，SNR）损失严重的情况，提出了一种结合卷积神经网络（Convolutional Neural Network，CNN）与Transformer并联的混合注意力的频谱感知网络模型（Hybrid Attention Spectrum Sensing Network，HA-SenseNet），并且引入MixMatch半监督学习方法，通过有效处理局部特征和全局信息，显著降低模型对标签数据的依赖性。进一步地，为了帮助模型更加聚焦于频谱能量图像中信息丰富的区域，在CNN分支中设计高效频谱注意力模块（Efficient Spectral Attention Module，ESAM）以动态调整模型权重，有效解决了低SNR条件下频谱观测图像的视觉模糊和特征混淆问题。仿真结果表明，与其他分类模型和半监督学习方法相比，HA-SenseNet通过并联架构实现了显著的性能增益，尤其在-20~5 dB的SNR区间提升幅度达16.8%；在标签数据占比10%的稀缺场景下，分类精度接近全监督学习水平。 |
| 关键词: 频谱感知 标签稀缺 半监督学习 注意力机制 |
| DOI：10.20079/j.issn.1001-893x.241213002 |
|
| 基金项目:国家自然科学基金资助项目(U23A20279) |
|
| Radiation-source-driven Spectral State Sensing under Label Scarcity and Low SNR Conditions |
| WANG Zixin,WANG Xin,SHEN Bin |
| (School of Communications and Information Engineering，Chongqing University of Posts and Telecommunications，Chongqing 400065，China) |
| Abstract: |
| To address the scarcity of spectral data labels and the serious loss of signal-to-noise ratio(SNR) during signal transmission in complex wireless propagation environments，a hybrid attention spectrum sensing network(HA-SenseNet) model is proposed，which combines convolutional neural network(CNN) and Transformer in parallel and introduces the MixMatch semi-supervised learning method to significantly reduce the model’s dependence on labeled data by effectively processing local features and global information.Further，in order to help the model focus more on the information-rich regions in the spectral energy image，the efficient spectral attention module(ESAM) is designed in the CNN branch to dynamically adjust the model weights，which effectively solves the problems of visual blurring and feature confusion of the spectral observation image under the low SNR condition.Simulation results show that compared with other classification models and semi-supervised learning methods，HA-SenseNet achieves a significant performance gain through the concatenated architecture，especially in the SNR interval from -20 dB to 5dB with an improvement of 16.8%；in the sparse scenario where the labeled data accounts for 10% of the data，the classification accuracy is close to the level of the fully-supervised learning. |
| Key words: spectrum sensing label scarcity semi-supervised learning attention mechanism |
《电讯技术》编辑部 