基于图最小割聚类注意力的自动调制分类网络

A Graph MinCut-Clustering Attention Based Automatic Modulation Classification Network

摘要:

面向智能频谱监测场景下动态电磁环境中自动调制分类复杂度高、低信噪比鲁棒性不足的问题，研究了一种基于图最小割聚类注意力的自动调制分类方法。采用I/Q双通道时序投影与旋转增强构造信号节点，利用数据增强和特征提取方法得到符号级节点特征，并通过可学习邻接矩阵、图粗化和动态图注意力聚合联合建模信号拓扑关系。以RadioML2016.10a数据集开展测试，在测试信噪比范围内平均识别准确率较基线方法提升，训练参数量和浮点运算次数均较在低水平。结果表明，该方法在保持较高识别精度的同时有效降低了模型复杂度，可满足智能频谱监测场景下自动调制分类的轻量化部署需求。

Abstract:

For automatic modulation classification in intelligent spectrum monitoring under dynamic electromagnetic environments, where high model complexity and insufficient robustness at low signal-to-noise ratios remain challenging, a graph MinCut-clustering attention based method was investigated. Signal nodes were constructed by using I/Q dual-channel temporal projection and rotation augmentation, symbol-level node features were obtained through data augmentation and feature extraction, and signal topological relationships were jointly modeled through a learnable adjacency matrix, graph coarsening, and dynamic graph attention aggregation. Tests were conducted on the RadioML2016.10a dataset, and the average recognition accuracy within the tested signal-to-noise ratio range was improved over baseline methods, while both the number of trainable parameters and floating-point operations remained low. The results show that the proposed method effectively reduces model complexity while maintaining high recognition accuracy, meeting the requirement for lightweight deployment of automatic modulation classification in intelligent spectrum monitoring scenarios.