Design Of Physical Layer Receiver Model Based On Deep Learning

For traditional communication systems,there are a large number of algorithms have been optimized for each communication module.Although the optimization can improve the performance of the single module,it does not mean that the performance of the whole system is optimal.Secondly,the existing communication system can only work in pre-optimized mode,and cannot perform the optimization automatically with the change of the demands or the environment,which is lack of certain flexibility.Deep learning is structured and intelligent.By applying deep learning to communication system,it can better adapt to the next generation communication system requirements or some specific application scenarios.The receiver is an important part of the communication system.Through deep learning,the receiver can be optimized as a whole to improve the performance of the communication system,including the channel estimation,channel equalization,demodulation and decoding processes.In order to study the feasibility and performance of deep learning in the communication physical layer,the feasibility of using the neural network to realize the basic demodulator module of the communication system is firstly analyzed.A simple neural network demodulator and a demodulator based on improved K-means algorithm are implemented.Compared with the traditional hard decision demodulator in the Gauss channel,the neural network-based demodulator performs better in the bit error rate with the training in different modulation modes.Meanwhile,the K-means based demodulator performs relatively high in the bit error rate with the less requirement of the training data.On this basis,the physical layer receiver is realized with three neural networks.Consisting of the frequency offset correction network,the channel estimation network and the channel equalization and demodulator network,the receiver is optimized as a whole.The simulation results show that the bit error rate of the neural network-based receiver is lower than that of the traditional receiver in the fading channel.While wireless communication often works in a fading environment,the introduction of deep learning in physical layer helps to improve the performance of communication systems.In order to further explore the feasibility of the application of deep learning in the physical layer of the actual communication system,by taking the OFDM(Orthogonal Frequency Division Multiplexing)system as an example,an OFDM receiver is designed based on deep learning.Firstly,the performance is analyzed by training and processing the channel estimation,channel equalization and demodulator with single neural network models.The simulation results show that neural network-based receiver performs better than the OFDM receiver based on the MMSE(Minimum Mean Square Error)and the LS(Least Square)channel estimation algorithms when the number of pilots is small.With the increasing of the pilots,the neural network-based OFDM receiver still performs better than the LS algorithm.While compared to the MMSE channel estimation algorithm,the neural network-based OFDM receiver performs relatively poor in the case of high SNR.Therefore,combining the neural network with the traditional LS algorithm and ZF(Zero Forcing)algorithm,a neural network receiver based on LS assistance is designed.The simulation results show that the improved neural network receiver performs better in bit error rate performance,but the training complexity is also high.