Robust deep learning based channel estimation technique for OFDM system

Abstract

Wireless communication channels have characteristics such as time variation, multipath, time delay, and Doppler frequency shift. Orthogonal frequency divi- sion multiplexing, as a multi-carrier digital modulation technology, can convert frequency-selective fading channels into a series of parallel narrow-band flat fad- ing channels. It has the advantages of strong multi-path resistance, high spec- trum utilization, and simple implementation. Facing the complex and changeable communication environment, channel estimation is a necessary means to ensure reliable communication. The main problem faced in channel estimation methods is to ensure the effect of channel estimation while also considering the com- plexity of the estimation method in a time-varying environment. As an emerg- ing data analysis and processing method, deep learning has unique advantages in learning the inherent laws of data and the input-output mapping relation- ship. This article aims at the channel estimation problem in OFDM systems and plans to explore channel estimation combined with deep learning. In order to directly learn the nonlinear mapping relationship between input and output in the wireless communication system and fit the channel characteristics, this dissertation combines the residual structure with the deep neural network and proposes an improved model of DNN, Res-DNN. Res-DNN replaces the 3 hidden layers in DNN with 3 residual blocks, and learns the non-linear mapping relationship between the transmitted and received signals by using the received data signals and the guide frequency signals at the receiving end as the model input data and the known signals at the transmitting end as the labels. Ex- perimental results show that when the number of pilots is small or there is no guard interval between OFDM symbols, the channel estimation method based on the Res-DNN model can still maintain high estimation accuracy and has good robustness.