Integrating Deep Learning for Hardware Impairments Mitigation in Communication Systems

Doktorsavhandling, 2024

The rapid growth of wireless communication, especially with the deployment of massive multiple-input multiple-output (MIMO) systems, has driven up power consumption, requiring more cost-efficient hardware. However, lower-cost components often introduce impairments such as power amplifier (PA) nonlinearity, phase noise (PN), in-phase and quadrature (IQ) imbalance, and mutual coupling, leading to degraded system performance. Classical model-based methods face challenges in effectively mitigating these impairments while maintaining performance and low computational complexity. In contrast, deep learning techniques offer a promising alternative by providing more adaptable and efficient solutions. This thesis explores deep learning-based methods for mitigating hardware impairments in communication systems, aiming to enhance both performance and power efficiency.
We first focus on single-input single-output (SISO) systems. In Paper A, we address the joint mitigation of IQ imbalance and PA nonlinearity using digital predistortion (DPD). Classical model-based methods often underperform in joint mitigation, and existing neural network (NN)-based methods are computationally demanding. To resolve these issues, we propose a novel NN-based DPD model combined with an NN pruning technique. This approach provides a more power-efficient solution for mitigating the combined impairments of the PA and IQ modulator compared to existing models. In Paper B, we further explore training DPD with low-sampling rate data. Supervised learning methods rely on high sampling-rate feedback paths, which are costly in wideband and multi-antenna scenarios. To overcome this, we introduce a reinforcement learning (RL)-based DPD learning algorithm that reduces the reliance on such high-sampling rate feedback paths while maintaining effective learning, making the DPD optimization more power-efficient.
We next shift our focus to mitigation techniques in massive MIMO systems. In Paper C, we tackle the joint mitigation of PA linearization and antenna crosstalk in massive multi-user (MU) MIMO orthogonal frequency division multiplexing (OFDM) networks. In such systems, the large number of antennas and corresponding PAs significantly increase the computational complexity of conventional DPD. While deploying DPD per user equipment (UE) in the frequency-domain (FD) instead of per PA in the time-domain (TD) can reduce complexity, the literature lacks proper FD DPD models. To address this, we propose a low-complexity FD convolutional neural network (CNN)-based DPD model. This model is effective in a line-of-sight (LOS) channel with fewer UEs.
Finally, Paper D explores joint channel and PN estimation in cell-free massive MIMO OFDM systems. Several previous studies assume single-carrier PN models to OFDM systems, leading to mismatches and overly optimistic performance predictions. We consider two setups: shared and separate local oscillators (LOs) between distributed access points (APs), which introduce uncorrelated and correlated PN respectively. We propose novel distributed and centralized joint PN and channel estimators, including a deep learning-based channel estimator, which demonstrates improved performance in both PN and channel estimation.
To summarize, this thesis explores the implementation of various deep learning-based techniques to effectively mitigate hardware impairments across different communication systems.