Design Of OFDM Systems Based On Convolutional Neural Network

The "14th Five-Year Plan" clearly pointed out the key focus on the new generation of artificial intelligence in the frontier field of science and technology.At the same time,deep learning is becoming more brilliant in life services such as intelligent transportation,culture and entertainment,and medical health.It has prompted scholars in the communication field to conduct extensive research on deep learning and use neural networks to optimize the design of traditional communication systems.To this end,around the realization of Orthogonal Frequency Division Multiplexing(OFDM)system,which is one of the key technology in the physical layer of wireless communications,this article focuse on designing of the OFDM system,which based on neural networks.The work flow and research results are as follows:Aiming to improve the reliability and effectiveness for the Physical Layer of the wireless communication systems,an Orthogonal Frequency Division Multiplexing Autoencoder(AE)transmission scheme AE-OFDM is proposed,which based on Convolutional Neural Network(CNN).This scheme using CNN structure to achieves the best transmission scheme under various complex channel environments through the end-to-end learning algorithm,which also has a deeper network structure and lower the network parameters.The simulation results have shown that in different wireless channel environments such as Additive Gaussian,hybrid impulse,and multipath channels,the proposed AE-OFDM system has good generalization ability and supports different band-utilization efficiency and coding efficiency.Besides,it has better block error rate performance is better than the expert OFDM systems in the higher band-utilization efficiency system.Aiming to improving the accuracy of the channel estimator in the OFDM system,we propose a deep learning Super-Resolution Reconstruction based OFDM channel estimator SR-CE,which used Dense Convolutional Neural Networks(Dense-Net)with dense connections and feature multiplexing to reconstruct low-resolution pilot images into high-resolution full Channel Frequency Responses(CFRs)images.The simulation results have shown that SR-CE can estimate the perfect CFRs well in the slow fading.In the fast fading,it has a higher channel estimation accuracy and achieved with fewer neural network parameters,thereby alleviates the floor effect of BLER performance.In order to further optimize the channel estimation scheme of SR-CE,an unsupervised deep learning Channel Estimation scheme UL-CE is proposed.This CE algorithm introduces the unsupervised deep learning method to estimation channel information,which can perform model training in the case of unknown CFRs.Also,we propose a double Elu activation function with positive and negative polarity,and minimize the Manhattan distance as the destination function to reduce the complexity of the neural network and improve the accuracy of channel estimation.The simulation results have shown that in multipath fading channels,the proposed CE algorithm has lower complexity,faster iterative convergence speed,and more accurate channel estimation results than the currently channel estimation algorithm which based on deep learning.