基于模型驱动深度学习的OTFS检测方法  

Signal Detection Based on Model-driven Deep Learning for OTFS

作　　者：魏新龙 李立[1] 温驰 靳一[1] 徐常志 WEI Xinlong;LI Li;WEN Chi;JIN Yi;XU Changzhi(China Academy of Space Technology(Xi’an),Xi’an 710100,China)

机构地区：[1]中国空间技术研究院西安分院,西安710100

出　　处：《电讯技术》2024年第8期1181-1186,共6页Telecommunication Engineering

基　　金：国家自然科学基金资助项目(61801377,62001375);国家重点研发计划(2019YFB1803102)。

摘　　要：正交时频空(Orthogonal Time Frequency Space, OTFS)调制技术凭借对多普勒频移的优良抗性,保证了高动态场景下的可靠性通信。与大多数OTFS信号检测方案相比,基于深度学习(Deep Learning, DL)的OTFS检测器不需要耗费高额的导频能量,以此获得精确的信道状态信息。基于多维输入的卷积神经网络(Convolutional Neural Networks, CNN)和一维输入的深度神经网络(Deep Neural Networks, DNN),搭建了OTFS信号检测模型,并结合OTFS的输入输出关系,以模型驱动,提出一种部分输入方法。与数据驱动DL相比,该方法沿时延轴截断输入数据,仅向网络输入与待检测信号相关性强的部分接收信号。该方法不仅减小了数据驱动CNN和DNN的训练参数量,降低了训练复杂度,而且检测性能也不弱于传统的线性最小均方误差(Linear Minimum Mean Square Error, LMMSE)算法。Orthogonal time frequency space(OTFS)modulation provides reliable communications in highmobility scenarios due to its good robustness against high Doppler shifts.Compared with most of the OTFS signal detection schemes,the OTFS detector based on deep learning(DL)does not cost high pilot power to get the perfect channel state information.In this paper,the OTFS detector is modeled using convolutional neural networks(CNN)with multi-dimensional input and deep neural networks(DNN)with onedimensional input,respectively.Furthermore,according to the input-output relation for OTFS systems,a novel model-driven part-input scheme is presented.Compared with the data-driven DL,the presented scheme truncates input data along the delay axis and requires part of the received signal,which has a strong correlation with the transmitted signal to be detected.The proposed scheme has fewer trainable parameters and lower complexity than the data-driven CNN and DNN,and is similar to or better than the conventional linear minimum mean square error(LMMSE)algorithm in terms of performance.

关 键 词：正交时频空(OTFS) 信号检测 深度学习 卷积神经网络 深度神经网络 

分 类 号：TN929[电子电信—通信与信息系统]