一种基于深度学习的OFDM信号检测方法  

An OFDM signal detection method based on deep learning

作　　者：高俊伟 郭晓冉 禹永植[1] GAO Junwei;GUO Xiaoran;YU Yongzhi(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China;32181 Troops of PLA,Shijiazhuang 050000,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]32181部队,河北石家庄050000

出　　处：《应用科技》2021年第3期68-72,共5页Applied Science and Technology

摘　　要：针对传统算法在正交频分复用(OFDM)系统导频数量较少时误符号率较高的问题,提出一种基于深度学习的OFDM信号检测方法,该算法设计一种信号检测网络,其信号检测网络可以代替传统算法中的信道估计和均衡,主要包含:输入层、双向长短记忆神经网络(Bidirectional long short memory neural network,BiLSTM)层、全连接层、softmax层以及分类层。首先需要构建BiLSTM,然后利用3GPP TR38.901信道模型下生成的数据对已经构建好的神经网络进行训练,最后则可将训练好的神经网络应用于OFDM系统之中,对整个系统进行信号检测。仿真结果表明,BiLSTM信号检测网络采用的是端到端的方式进行OFDM信号检测,与传统的信号检测方法相比,BiLSTM信号检测网络在误符号率为10^(−3)时,有5~6 dB的性能提升,与同类型的采用端到端的LSTM信号检测网络的算法相比,该算法在误符号率为10^(−3)时,有2~3 dB的性能提升。Aiming at the problem of high symbol error rate in the traditional algorithm when the number of pilots in the OFDM system is small,an OFDM(Orthogonal Frequency Division Multiple)signal detection method based on deep learning is proposed.The proposed algorithm is mainly to design a signal detection network.It can replace channel estimation and equalization in traditional algorithms.The network mainly includes input layer,BiLSTM(Bidirectional long short memory neural network)layer,fully connected layer,softmax layer and classification layer.First,it is necessary to build BiLSTM,and then using the data generated under the 3GPP TR38.901 channel model to train the built neural network.Finally,the trained neural network can be applied to the OFDM system.Signal detection is performed on the entire system.Simulation results show that the algorithm proposed in this paper uses an end-to-end approach for OFDM signal detection.Compared with traditional signal detection methods,BiLSTM signal detection network has 5-6 dB performance improvement when the symbol error rate is 10^(-3).Compared with the same type of algorithm using end-to-end LSTM signal detection network,BiLSTM signal detection network has 2-3 dB performance improvement when the symbol error rate is 10^(-3).

关 键 词：正交频分复用 BiLSTM信号检测网络 信号检测 误符号率 神经网络 端到端 导频 深度学习 

分 类 号：TN929.5[电子电信—通信与信息系统]