Abstract:

To address the limited accuracy and high complexity of channel estimation in Orthogonal Frequency Division Multiplexing （OFDM） systems， a Deep Neural Network （DNN） based channel estimation algorithm for OFDM systems was proposed. First， the channel estimation and equalization processes used in traditional communication systems were replaced by the proposed algorithm， and the pilot information and data blocks at the receiver were used as input features， enabling rapid learning and estimation of channel characteristics. Second， during the model training phase， OFDM frames containing pilot blocks and data blocks were constructed， and combined with a wireless channel propagation model， a large number of simulated data pairs were generated， thereby ensuring that the training data covered a variety of complex channel environments and noise conditions. Finally， the OFDM system was used to perform a nonlinear mapping of the input-output relationship， transforming this mapping relationship into a task suitable for supervised DNN learning， thereby enabling the trained model to output accurate channel estimation results at the receiver efficiently， as well as recovering the original transmit data reliably and improving overall system performance. At the same time， a key computational module for DNN model training was designed on Field Programmable Gate Array （FPGA） to monitor and optimize the trained model in real time. Experimental results showed that， compared with the Least Squares （LS）， Minimum Mean Square Error （MMSE）， and Discrete Fourier Transform-Least Squares-Wiener （DFT-LS-WIENER） algorithms， at a Signal-to-Noise Ratio （SNR） of 10 dB， the proposed algorithm reduced the Bit Error Rate （BER） by 92.82%， 80.32%， and 72.27%， respectively， and outperformed the above comparative algorithms in Normalized Mean Squared Error （NMSE）. The proposed algorithm has good anti-interference performance and estimation accuracy under complex channel conditions； furthermore， it has moderate resource consumption on an FPGA hardware platform， and controllable operating power consumption， making it suitable for deployment on FPGA.

Key words: Deep Neural Network (DNN), channel estimation, Orthogonal Frequency Division Multiplexing (OFDM), wireless channel, Field Programmable Gate Array (FPGA)

摘要：

针对正交频分复用（OFDM）系统中信道估计精度受限及复杂度较高的问题，提出一种基于深度神经网络（DNN）的OFDM系统信道估计算法。首先，利用所提DNN模型替代传统通信系统中依赖的信道估计与均衡过程，并以接收端的导频信息和数据块作为输入特征，实现对信道特性的快速学习与估计；其次，在模型训练阶段，构建包含导频块和数据块的OFDM帧，并结合无线信道传播模型生成大规模的仿真数据对，以保证训练数据覆盖多种复杂信道环境和噪声条件；最后，利用OFDM系统完成对输入输出关系的非线性映射，并将该映射关系转化为适合DNN监督学习的任务，从而使训练完成的模型能在接收端高效输出精确的信道估计结果，且实现对原始传输数据的可靠恢复并提升系统的整体性能。同时，基于现场可编程门阵列（FPGA）设计了DNN模型训练的主要计算模块，实现对训练模型的实时监控与优化。实验结果表明，与最小二乘（LS）算法、最小均方误差（MMSE）算法以及DFT-LS-WIENER算法相比，在信噪比（SNR）为10 dB时，本文算法分别将误码率（BER）降低了92.82%、80.32%、72.27%，归一化均方误差（NMSE）也小于对比算法。本文算法在复杂信道条件下具有良好的抗干扰性能和估计精度；而且在FPGA硬件平台上的资源消耗适中，运行功耗处于可控范围内，可部署在FPGA上。

关键词: 深度神经网络, 信道估计, 正交频分复用, 无线信道, 现场可编程门阵列