| 摘要: |
| 针对基于机器学习的无线电地图（Radio Map，RM）重构方法在测量样本稀疏且分布非均匀时存在过拟合及精度不足问题，提出了一种基于卷积块注意力模块与亲和传播聚类U型网络（Convolutional Block Attention Module and Affinity Propagation Clustering Based UNet，CAP-UNet）的RM重构方法。首先，利用亲和传播聚类提取稀疏样本的空间结构先验，作为补充特征输入U型网络，以缓解样本非均匀分布引起的信息缺失；其次，通过在U型网络的跳跃连接处嵌入卷积块注意力模块，增强模型对建筑物边缘等关键传播特征的聚焦能力；最后，设计双网络交叉训练策略，通过迭代协同进一步优化重构性能。在样本比例1%、失衡比例为100%的非均匀稀疏条件下比较了CAP-UNet与其他基于UNet重构算法的精度，结果表明，CAP-UNet归一化均方误差较RME-GAN、RadioUNet分别降低了53%、24.6%。 |
| 关键词: 无线电地图重构 U型网络 亲和传播聚类(APC) 交叉训练 |
| DOI：10.20079/j.issn.1001-893x.260129001 |
|
| 基金项目:通信抗干扰全国重点实验室基金项目（2024-kgr-JJ-08）；国防科技大学自主科研基金项目(ZK24-58) |
|
| A Cross Training Enhanced UNet for Precise Radio Map Reconstruction |
| WANG Ran,ZHANG Yusi,PAN Yijin,ZHAO Liping,ZHANG Xiaobao,LI Xinran |
| (1.National Mobile Communications Research Laboratory，School of Information Science and Engineering，Southeast University，Nanjing 210007，China; 2.The 63rd Research Institute，National University of Defense Technology，Nanjing 210007，China; 3.National Key Laboratory of Test&Evaluation for ElectroMagnetic Space Security，Nanjing 210007，China;4.A Unit of the PLA Army，Xuzhou 221000，China) |
| Abstract: |
| To address overfitting and insufficient accuracy in machine-learning-based radio maps (RMs) reconstruction under sparse and non-uniformly distributed samples，a UNet-based architecture integrated with a convolutional block attention module and affinity propagation clustering (CAP-UNet) is proposed.First，affinity propagation clustering is employed to extract spatial structural priors from sparse samples，which are fed into the network as complementary features.Second，a convolutional block attention module is embedded into the skip connections of the UNet to enhance the ability to focus on key propagation features.Furthermore，a dual-network cross-training strategy is designed to improve the reconstruction performance through iterative collaborative optimization.Simulation results show that，under a sampling ratio of 1% and an imbalance ratio of 100%，CAP-UNet reduces the normalized mean square error by 53% and 24.6% compared with RME-GAN and RadioUNet，respectively. |
| Key words: radio maps reconstruction UNet affinity propagation clustering(APC) cross-training |
《电讯技术》编辑部 