基于深度学习的分数多普勒信道估计技术  

Fractional Doppler channel estimation technology based on deep learning

作　　者：蒲旭敏 王可豪 陈伟聪 刘雁翔 陈前斌[1] PU Xumin;WANG Kehao;CHEN Weicong;LIU Yanxiang;CHEN Qianbin(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;National Mobile Communications Research Laboratory,Southeast University,Nanjing 211189,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]东南大学移动通信国家重点实验室,南京211189

出　　处：《东南大学学报(自然科学版)》2025年第2期593-602,共10页Journal of Southeast University:Natural Science Edition

基　　金：国家自然科学基金联合基金资助项目(U23A20279);国家自然科学基金资助项目(62401137);江苏省自然科学基金资助项目(BK20241281).

摘　　要：针对正交时频空间系统在整数多普勒模型中多普勒分辨率较低,不适用于实际通信场景的问题,在多输入多输出的正交时频空间(multiple⁃input multiple⁃output orthogonal time frequency space,MIMO⁃OTFS)调制系统中考虑分数多普勒信道模型,可有效提升多普勒频移分辨率,但同时会产生虚拟路径,导致多普勒间干扰。因此研究了一种模型驱动的学习去噪近似消息传递(learned denoising⁃based approximate message passing,LDAMP)算法,对含有多普勒间干扰的分数多普勒信道进行估计。该算法以去噪近似消息传递(denoising⁃based approximate message passing,DAMP)算法为基础,构建了一个可解释的神经网络框架,并选用去噪卷积神经网络(denoising convolutional neural network,DnCNN)替代DAMP中的传统去噪器,通过学习噪声特征将之有效去除,进而显著提升后续信号处理性能。仿真结果表明,模型驱动的LDAMP算法结合了迭代算法的模型优势和深度学习强大的泛化能力,相较于传统算法,能够有效补偿多普勒间干扰带来的性能损失,实现更高的信道估计精度。The Doppler resolution of the orthogonal time frequency space system is low in the integer Doppler model,which is not suitable for the actual communication scenes.Fractional Doppler channel model was con⁃sidered in the multi⁃input multi⁃output orthogonal time frequency space(MIMO⁃OTFS)modulation system,which can effectively improve the Doppler shift resolution.However,it will produce virtual paths,resulting in inter⁃Doppler interference.Therefore,a model⁃driven learned denoising⁃based approximate message pass⁃ing(LDAMP)algorithm was studied to estimate the fractional Doppler channel with inter⁃Doppler interfer⁃ence.The algorithm is based on denoising⁃based approximate message passing(DAMP)algorithm,and con⁃structs an interpretable neural network framework.Denoising convolutional neural network(DnCNN)was chosen to replace the traditional denoiser in DAMP,which can effectively remove the noise by learning the noise features,and then significantly improve the processing performance of subsequent signal.Simulation re⁃sults show that the model⁃driven LDAMP algorithm combines the model advantage of iterative algorithm and the powerful generalization capability of deep learning,and can effectively compensate the performance loss caused by inter⁃Doppler interference and achieve higher channel estimation accuracy compared to the tradi⁃tional algorithm.

关 键 词：正交时频空间 深度学习 分数多普勒 模型驱动 信道估计 

分 类 号：TN911.72[电子电信—通信与信息系统]