基於深度展開之壓縮通道估測用於智慧反射面板輔助之無線通訊

Abstract

為了滿足第五代（Fifth Generation, 5G）無線通信系統的性能要求，人們已經開發了各種相關技術，如大規模多輸入多輸出（Massive Multiple-Input Multiple-Output, Massive MIMO）和毫米波（Millimeter-Wave, mmWave）通訊。然而，實現這些技術需要增加能源消耗和硬體成本，這可能不是未來第六代（Sixth Generation, 6G）無線網路的可擴展解決方案。近年來，智慧反射面板（Intelligent Reflecting Surface, IRS）已經成為具發展前景的技術，能夠以低成本高效益的方式增強頻譜效率和能源效率。透過智能地調控低成本的反射元件，IRS可以將入射信號反射到指定的方向，以提高通道容量和可靠性。然而，由於大量低成本的被動反射元件，並不具備信號感知和信號處理能力，因此通道狀態資訊（Channel State Information, CSI）的獲取需要大量的訓練負擔（Training Overhead）。 儘管利用壓縮感知（Compressive Sensing, CS）的技術可以減少訓練負擔，但傳統的CS算法在低訓練負擔下無法達到滿意的通道估測效能。在本論文中，我們利用深度展開（Deep Unfolding）技術來結合壓縮感知和深度學習（Deep Learning, DL）的優點。首先，我們提出了一種新型的超網絡輔助學習近似訊息傳遞（Learned Approximate Message Passing, LAMP）網絡，以資料驅動的方式提高通道估計的準確性。此外，我們提出了一種具有列壓縮的雙階段LAMP網絡（RCTS-LAMP），通過分解通道估測過程，實現了更好的復雜度和準確度之間的平衡。此外，RCTS-LAMP網絡還可以應用於直接通道（Direct Channel）和級聯通道（Cascaded Channel）的聯合估測，而無需額外的網絡。最後，通過從多測量向量（Multiple-Measurement Vector, MMV）角度解決多用戶通道估測問題，我們提出的RCTS-LAMP-MMV網絡可以顯著降低IRS輔助多用戶系統的訓練負擔。

To fulfill the performance requirements of the fifth-generation (5G) wireless communication system, a variety of enabling technologies have been developed, such as massive multiple-input multiple-output (MIMO) and millimeter-wave (mmWave) communication. However, implementing these techniques requires increasing energy consumption and hardware cost, which may not be a scalable solution for the sixth-generation (6G) wireless networks. Recently, intelligent reflecting surface (IRS) has emerged as a promising technique to enhance spectral and energy efficiency cost-effectively. By smartly reconfiguring the low-cost reflective elements, IRS can reflect the incident waves toward the desired directions to improve communication capacity and reliability. However, the acquisition of channel state information (CSI) requires large training overhead due to the large number of low-cost passive reflective elements that do not have the capabilities of sensing and signal processing. Although the training overhead can be reduced with compressive sensing (CS) techniques, conventional CS algorithms cannot achieve satisfactory estimation performance under low training overhead. In this dissertation, we leverage the deep unfolding technique to combine the advantages of CS and deep learning (DL). First, we present a novel hypernetwork-assisted learned approximate message passing (LAMP) network to improve the channel estimation accuracy in a data-driven manner. Furthermore, we propose a two-stage LAMP network with row compression (RCTS-LAMP), which achieves a better trade-off between complexity and accuracy by decomposing the estimation process. Besides, the RCTS-LAMP network can be applied to the joint estimation of direct and cascaded channels without an extra network. Finally, by solving the multi-user channel estimation problem from the multiple-measurement vector (MMV) perspective, the proposed RCTS-LAMP-MMV network can significantly reduce the training overhead for IRS-aided multi-user systems.