The massive multiple-input multiple-output (MIMO) system is widely recognized as a key enabler for the fifth-generation (5G) and future wireless communication systems. For massive MIMO systems, estimation of the channel state information (CSI) is essential for high spectral efficiency designs. However, accurate channel estimation (CE) is difficult for massive MIMO systems because the number of channel parameters to estimate is huge due to the large number of antennas. In Part I of the thesis, we review some traditional channel estimation methods, including the state-of-the-art compressive sensing-based solutions exploiting different channel sparsity structures. We also propose a compressive sensing-based solution for joint channel estimation and localization for massive MIMO systems wit...[ Read more ]

The massive multiple-input multiple-output (MIMO) system is widely recognized as a key enabler for the fifth-generation (5G) and future wireless communication systems. For massive MIMO systems, estimation of the channel state information (CSI) is essential for high spectral efficiency designs. However, accurate channel estimation (CE) is difficult for massive MIMO systems because the number of channel parameters to estimate is huge due to the large number of antennas. In Part I of the thesis, we review some traditional channel estimation methods, including the state-of-the-art compressive sensing-based solutions exploiting different channel sparsity structures. We also propose a compressive sensing-based solution for joint channel estimation and localization for massive MIMO systems with the presence of phase noise. However, these compressive sensing-based solutions usually need a lot of iterations and require computational intensive modules in each step, making it difficult to apply these iterative algorithms for real-time CE in 5G networks with low-latency requirements.

The primary advantages of deep learning (DL)-based CE is fast implementation and its data-driven nature. However, one of the major limitations of existing DL-based solution is the requirement of offline training before they can be used online for channel inferencing. As such, the offline trained deep neural network (DNN) cannot adapt to the model mismatches when the channel model used to generate the offline training labels is different with that in the actual scenarios. In Part II of the thesis, we propose an online training framework for DNN-based CE, which can simultaneously perform channel inferencing as well as update of the DNN weights without the need of true channel labels. To realize this, we propose a few axioms for a legitimate online loss function, based on which we develop an online training algorithm with error analysis for massive MIMO CE. The online training framework is then extended to another two scenarios. In the first extension, we consider online DL-based CE for massive MIMO system with nonlinear amplifier distortions, where we propose an online loss function incorporating the unknown nonlinearity. Based on this, a two-stage DNN architecture is proposed to facilitate the joint training of the channel estimator and the nonlinear modules. In the second extension, we propose a federated online training algorithm for DNN-based CE in multi-user (MU) massive MIMO systems, where the designed two-tier DNN can learn and utilize the partial common sparsity structure in the MU MIMO channels. The CSI at the transmitter (CSIT) and the CSI at the receiver (CSIR) can be simultaneously generated in real-time at the BS and at each UE, respectively, in the federated online training algorithm. It is shown from extensive simulations that the proposed online training framework achieves superb CE accuracy with much faster channel inferencing over the state-of-the-art methods, and is able to adapt to various channel model mismatches.

In Part III of the thesis, in order to further reduce the feedback overhead, we propose a Variational Bayesian Autoencoder (VBA) solution for nonlinear compression of the CSI and feedback with quantization in massive MIMO systems. In contrast with existing DL-based schemes, the outputs of the autoencoder/decoder are probability distributions, which enables us to incorporate the model-assisted knowledge of low-dimensional latent space and the sparsity of channel in the training formulation. Based on this, we propose a network architecture, called CsiVBA, which is able to automatically learn a sparse representation in the latent space and recover the CSI exploiting the angular-delay domain sparsity. A multi-user VBA network is also proposed to explore the partial common sparsity in the MU MIMO channels and facilitates the learning of support information in the feedback features to further reduce the feedback overhead. Simulation results show that the proposed scheme achieves better rate-distortion trade-offs than the state-of-the-art solutions for a wide range of compression rates.