| In recent years,mobile communication technology has rapidly developed and become one of the most important fields.Fifth Generation(5G)technology has become the fundamental interconnection network that supports various industries.With the continuous expansion of service coverage,the underlying technologies face increasingly high requirements,especially Orthogonal Frequency Division Multiplexing(OFDM)which is one of the key technologies in the physical layer.OFDM has attracted much attention from relevant scholars due to its high-speed data communication capabilities and robustness to multipath time delay extension.In wireless communication systems based on OFDM,accurate estimation of the Channel State Information(CSI)is required for any channel to improve the performance of the communication system.Therefore,accurate channel estimation becomes particularly important.The traditional channel estimation algorithm,Least Square(LS),is widely used in practice due to its simplicity,but it ignores the impact of channel noise,leading to poor estimation performance.The Minimum Mean Square Error(MMSE)method obtains better estimation performance at the channel response by obtaining channel and noise statistical information in advance.However,it requires solving the channel’s autocorrelation matrix,leading to high complexity.In recent years,deep learning has made significant strides in various fields.With the vigorous development of intelligent communication,the application of deep learning in channel estimation has gradually matured.In this paper,we use offline reinforcement learning to explore the hidden relationship between channel and noise behind the data to achieve accurate channel estimation.We analyze existing applications of deep learning in channel estimation and traditional channel estimation methods.The main contributions of this paper are as follows:This paper first analyzes the limitations of traditional channel estimation methods through simulation and proposes a super-resolution(SR)network-based Rayleigh channel estimation algorithm to address the problem of difficult-to-obtain channel prior information.This method models the channel response matrix obtained by LS as a two-dimensional image and utilizes image super-resolution and image restoration(IR)techniques for learning optimization to obtain the channel response at unknown symbols.Simulation results show that the proposed algorithm is significantly superior to the traditional linear minimum mean square error(LMMSE)method,with a 3d B gain compared to other SR-based channel estimation algorithms at low signal-to-noise ratio(SNR)and a 5d B gain at high SNR.Based on the research results mentioned above,this paper extends the application scenario to channel estimation in high-speed mobile environments.Firstly,the pre-processing of the dataset is optimized.Unlike the current mainstream approach,this paper uses the adjacent OFDM symbols to build a three-dimensional channel information matrix,preserving the relative positions of the real and imaginary parts,so that the deep learning model can fully learn the amplitude and phase information during the channel variation process.Then,a bidirectional long short-term memory network(Bi LSTM)is introduced to fully utilize the learned channel information for channel estimation.Simulation results from four aspects,including modulation order,mobility,frame length,and deep learning model architecture,demonstrate that this algorithm outperforms other existing deep learning algorithms for channel estimation. |
