| In an orthogonal frequency division multiplexing(Orthogonal Frequency Division Multiplexing,OFDM)communication system,the multipath effect of the wireless channel distorts and attenuates the signal during transmission,and the relative motion between the transmitter and receiver produces Doppler effect.The effect eventually causes the signal carrier to shift.Therefore,it is very important to understand the channel characteristics.In order to recover the original signal from the signal that has been contaminated by noise,the traditional OFDM channel estimation algorithm estimates the channel impulse response based on the pilot,and then restores the original signal by interpolation and equalization.This thesis attempts to use machine learning to perform channel estimation in OFDM systems to reduce the use of pilots,improve spectrum utilization,and improve the accuracy of channel estimation.The results prove that neural networks after deep learning can better restore the originally sent signal.This paper designs and implements the following neural network to perform channel estimation:(1)Based on the traditional Least Squares(LS)channel estimation algorithm,a channel estimator based on a three-layer back propagation(BP)neural network is proposed.Each BP neural network estimate the channel response at a single frequency.The simulation results show that the performance of the BP neural network is better than the LS algorithm,which is slightly insufficient compared with the Minimum Mean Square Error(MMSE)algorithm.Although the MMSE algorithm has better performance than the BP neural network,the complexity of the BP neural network when performing online estimation is lower than that of the MMSE algorithm,and the BP neural network does not require prior information of the channel for online estimation.(2)A GA-BP neural network is trained,which formed by combining BP neural network and Genetic Algorithm(GA).Simulation results show that under various modulation schemes and pilot symbols of various lengths,the GA-BP neural network has higher estimation performance and faster convergence speed than the unimproved BP neural network,and when the pilot symbols is reduced,it's performance is superior to the traditional LS algorithm,equivalent to the Minimum Mean Square Error(MMSE)algorithm.(3)A five-layer fully connected neural network is designed for channel estimation,processing wireless OFDM channels in an end-to-end manner.Unlike existing OFDM receivers,which estimate the channel state information first and then recovers the transmitted symbols,the fully connected neural network internally estimates the channel state information and directly restores the transmitted symbols.From the simulation results,the performance of the fully connected neural network is superior to the LS algorithm,and equivalent to the performance of Linear Minimum Mean Square Error(LMMSE).In addition,when fewer training pilots are used,cyclic prefixes are omitted,and nonlinear clipping noise is present,the fully connected neural network are more robust than traditional LS and LMMSE channel estimation algorithms. |
