Channel Estimation For MIMO-OFDM Systems Based On Deep Learning

This thesis studies the channel estimation problem in multiple-input multiple-output orthogonal frequency division multiplexing(MIMO-OFDM)system from two directions.The first is the channel estimation problem of MIMO-OFDM at pilot symbols in a rich scattering environment.The other is the double selective fading channel estimation problem in MIMO-OFDM when the device moves.First,in the MIMO-OFDM channel estimation problem at pilot symbols in a rich scattering environment,we find that the channel sparsity in the delay-angle domain is severely compromised in a rich scattering environment so that most existing compressed sensing based techniques can harvest a very limited gain(if any)in reducing the channel estimation overhead.We propose a learning-based Turbo message passing(LTMP)algorithm to solve this problem.Instead of utilizing channel sparsity,LTMP can effectively extract channel features through deep learning and utilize channel continuity in the frequency domain through block linear modeling(BLM).We used the 3rd generation partnership project clustered delay line channel model to evaluate LTMP performance in MIMOOFDM channels.Simulation results show that when the normalized mean square error of channel estimation is-20 dB,the proposed channel estimation method has more than 5dB power gain than the existing algorithms.The algorithm also shows strong robustness in various environments.Then,we consider the channel estimation problem on the non-pilot symbol in the doubly selected channel.A double symbols one-dimensional convolutional neural network(DS-1DCNN)based on the data pilot-aided channel estimation method is proposed in a symbol-by-symbol channel estimation scheme.Specifically,BLM improves the channel estimation results on a single OFDM symbol.At the same time,DS-1DCNN considers the correlation between OFDM symbols,inhibits the error propagation between symbols,and greatly reduces the number of parameters.When the Doppler shift is 800 Hz,the modulation order is 256 QAM,and the normalized mean square error of channel estimation is-11 dB,DS-1DCNN can bring 4 dB power gain relative to the feedforward neural network.In the frame-by-frame channel estimation scheme,we find that the existing channel estimation methods based on convolutional neural networks do not fully consider the structural characteristics of the channel in OFDM frames and the practical problems of channel super resolution recovery but blindly stack the convolutional layers.Based on this,we propose a multi-scale channel recovery network(MCRNet)based on up-and-down sampling.The network’s core is the attention mechanism and the interlaced up-and-down sampling operation throughout the whole network.Since the reliability of channel state information on different resource elements is not consistent,the attention mechanism can automatically learn the more useful features and suppress the noise.The up-and-down sampling operation can ensure that the channel features at different scales are fully integrated and complementary,and the noise can be discarded during sampling.When the Doppler shift is 800 Hz,and the normalized mean square error of channel estimation is-17 dB,the proposed scheme can bring a power gain of 2 dB compared with the current best scheme.