Research On Channel Estimation Algorithms For OFDM And Multi-Antenna Systems Based On Deep Learning

With the rapid development of key communication technologies,the structure of a communication system is more complex and the number of antennas configured is also increasing.At the same time,higher requirements are put forward for the performance of various aspects of a communication system,such as high throughput,high data rate and so on.Channel estimation is always a key part of a wireless communication,and its efficiency is critical to the communication system.Therefore,high-precision and efficient channel estimation algorithms have always been the focus of research in the field of communications.With the increasing improvement of deep learning(DL)related theories,DL has been widely used in various fields and has achieved excellent results.Meanwhile,with the continuous optimization of DL algorithms and the improvement of hardware devices,DL technology has gradually been applied to the field of communication.Therefore,using DL technology to improve the estimation efficiency of complex channels in communication systems has become an important research direction.First,Orthogonal Frequency Division Multiplexing(OFDM)technology can well support multiple access and has strong robustness in frequency selective fading environment,so it has been widely used in communication in the system.Most of the current DL-based OFDM channel estimation algorithms must be pre-trained with large datasets,which results in excessive training overhead and discrepancies between training and testing phases.In response to these problems,this thesis proposes a DL-based OFDM channel estimation algorithm that does not require offline training-Untrained Channel Estimation Network(UTCENet).In UTCENet,the channel matrix obtained by Least Square(LS)estimation is first modeled as a two-dimensional image containing noise;then gradient descent is used to dynamically fit the parameters of the neural network for each noisy image,that is,the implicit prior information is obtained through the parameterization of the neural network;finally,the high-precision channel image is reconstructed by using the obtained implicit prior information.Although the UTCENet algorithm does not require offline training,the estimated performance is guaranteed due to exploiting the correlation between OFDM subcarriers.Numerical results show that,compared with traditional methods and existing DL-based channel estimation methods that require training,the proposed UTCENet algorithm not only has a significant performance improvement in estimation accuracy and estimation accuracy,but also reduces training overhead.Then,4G and post-4G Multi-input Multi-output(MIMO)OFDM becomes the key technology.This thesis further applies the proposed UTCENet algorithm to the MIMO-OFDM system.The LS estimation of the channel matrix between each antenna pair is modeled as a noise image,which is then superimposed in the spatial dimension.At the same time,the UTCENet network model is extended from single channel to multi-channel,in order to better utilize the correlation between OFDM subcarriers to improve the accuracy of channel estimation.The numerical results show that,compared with other DL channel estimation algorithms,the UTCENet algorithm has faster convergence speed and higher estimation accuracy.The possibility of extending the UTCENet algorithm from low-dimensional signals to high-dimensional signals is verified.Finally,due to the development of antenna technology and millimeter wave(mm Wave)technology,millimeter wave massive MIMO has become one of the hot key technologies in wireless communication systems.The research on channel estimation algorithms in this thesis is further extended to millimeter wave massive MIMO-OFDM system.Because the system has sparse characteristics,the channel estimation problem of the system can be modeled as a multi-measurement vector problem,and then the channel estimation can be realized by using compressed sensing theory.However,the calculation process of the classical matching pursuit method is complicated.To solve this problem,this thesis uses a deep neural network(DNN)on the basis of compressed sensing,and proposes a channel estimation algorithm based on DNN(DNN-CES)to achieve end-to-end channel estimation.The experimental results show that the DNN-CES algorithm has lower normalized mean square error and higher bit error rate than the classical matching pursuit greedy method,and the calculation process is simpler.