Research On OFDM Signal Processing Technique Based On Deep Learning

Orthogonal frequency division multiplexing(OFDM),as a multi-carrier modulation technique,has been widely used in wireless communication,due to its high spectrum utilization,resistance to frequency selective fading,and easy implementations.However,the transmission OFDM signal is usually affected by multipath fading channel and transceiver hardware impairments,which causes severe signal distortion and deteriorates the performance of the receiver.Signal estimation and compensation algorithms of receivers in communication systems can effectively alleviate the signal distortion.Therefore,in order to guarantee the signal transmission quality and system performance,it is crucial to investigate the signal processing algorithm of the receivers for OFDM systems.In the dissertation,the signal processing of the OFDM system is focused,and deep learning is considered at the OFDM receiver.The in-phase/quadrature(I/Q)imbalance estimation and compensation,channel estimation and signal detection algorithms are studied,and signal processing algorithms based on deep learning are designed to improve system performance.The main research contents of the dissertation are as follows:(1)The baseband transceiver structure of the OFDM system with I/Q imbalance is researched,and the cause of the I/Q imbalance and the negative influence on the signal are analyzed.Besides,the basic formulas of the OFDM signal process are derived,and the basic theory of deep learning is introduced.(2)In order to solve the signal detection problem of OFDM system receiver,two detection algorithms based on deep learning are designed,including CNN-OFDM and ZFNet algorithms.The CNN-OFDM can replace the traditional signal processing module and directly process the received frequency domain signals to recover OFDM symbols.In addition,the CNN-OFDM,which is not sensitive to the existence of cyclic prefix,can also greatly improve system performance under the condition of fixed channel parameters.The ZFNet is combined with the traditional zero forcing(ZF)algorithm,and the data initially detected by the ZF is regraded as one of the inputs,in order to effectively extract the signal features and recover the data.Simulation results show that ZFNet can improve the detection performance of conventional ZF algorithm,and the BER curves of ZFNet are close to those obtained by the ideal minimum mean square error(MMSE)algorithm.(3)In order to improve the accuracy of I/Q imbalance or channel estimation,IQIest Net estimation algorithm and LSNet channel estimation algorithm are proposed to solve the problem of signal distortion caused by I/Q imbalance and multipath fading channel.Both IQIest Net and LSNet use traditional estimation algorithm for preliminary estimation,and then use neural network to effectively extract I/Q imbalance or channel features to achieve accurate estimation.Besides,IQIest Net has simple network structure with few training parameters and is easy to implement,which can effectively improve the estimation performance of traditional I/Q imbalance estimation algorithms.LSNet channel estimation algorithm uses convolutional neural network to estimate channel parameters,which improves the accuracy of channel estimation while avoiding the complicated computation of MMSE algorithm.(4)The IQIest Net,LSNet and ZFNet models are integrated into the traditional signal processing module of OFDM receiver,and a joint estimation and signal detection algorithm combined with traditional methods is proposed.The joint algorithm is verified to be effective in improving the error performance of the traditional algorithm through simulation.Finally,the network models proposed in the dissertation are compared in detail,and the characteristics and advantages of each model are analyzed.