Research On Deep Learning-based Channel Estimation For OFDM Systems

With the rapid development of wireless communication,5G is gradually growing into a largescale basic Internet that supports various industries in the whole society.The substantial expansion of its service scope poses many challenges for the underlying technology,especially for the crucial component of the physical layer — Orthogonal Frequency Division Multiplexing(OFDM).Recently,deep learning has attracted extensive attention due to its excellent performance in computer vision and natural language processing.Its strong universality also provides new development space for traditional communications.This paper conducts an in-depth study on the channel estimation for OFDM systems,and explores the possible application of deep learning in this field.According to the analysis of traditional estimation algorithms and existing deep learning application cases,this paper follows the model-driven approach to solve the problems of traditional algorithms,including high dependence on prior information,large pilot overhead resulting in the waste of spectrum resources,and high computational complexity.Firstly,this paper introduces the OFDM system model and explains the fundamental principle of its effective resistance to multipath effects.To better understand the influence of wireless channel on signal transmission,this paper briefly analyzes the characteristics of channel fading from the perspectives of large-scale fading and small-scale fading,and gives an intuitive explanation through the simulation of common channel modeling methods.Moreover,three traditional channel estimation algorithms,i.e.,LS,MMSE,and LMMSE,are derived by mathematics.In addition,this papper illustrates several typical deep neural networks in detail,including their internal structure,parameter updating process,and related optimization algorithms.The accuracy improvement of traditional channel estimation algorithms often relies on the statistical characteristics of channel and noise.To solve the restrictions caused by the difficulty of obtaining prior information,this paper proposes a rayleigh channel estimation algorithm based on image super-resolution(SR)network.In this algorithm,the initial channel response matrix obtained by the LS algorithm is regarded as a two-dimensional low-resolution image,and then input into the SR network for resolution improvement,which can be equivalent to the increase of estimation accuracy.The SR network is originally designed for image restoration,and now is used to construct the mapping between low and high-precision channel estimates in this paper.The simulation results show that the proposed algorithm significantly outperforms the traditional LMMSE algorithm,and has strong competitiveness in data transmission efficiency and spectrum utilization.In the comparison of two different SR networks,the deep learning method based on the deconvolutional network is more prominent.Based on above research results,this paper extends the application scenario to the doublyselective channel under the basis expansion model(BEM),and optimizes the data set preprocessing method and the deep neural network structure.The main function of data set preprocessing is to convert complex numbers into the format suitable for network calculation.A reasonable data set structure can promote the efficiency of the algorithm.Different from the existing deep learning-based methods,this paper retains the relative positions of the real and imaginary parts in the matrix,and supplements the data with the channel response of the adjacent OFDM symbols,so as to make full use of the amplitude and phase information during channel variation.Simulations prove that this preprocessing method helps the proposed algorithm to get extra performance gain.Considering the computational complexity,this paper designs a lightweight fully convolutional neural network with three layers.In this network,the input low-precision channel estimates are transformed into highprecision outputs after feature extraction,nonlinear mapping,and deconvolutional expanding.Compared with the LMMSE algorithm,the proposed algorithm achievies higher estimation accuracy with lower computational complexity,and still has superior performance with limited number of pilots.