請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91228
標題: | 智慧反射表面輔助毫米波通信:基於加性注意力輔助對抗自編碼器的波束成型設計 Intelligent Reflecting Surface-Assisted Millimeter Wave Communications: Additive Attention-aided Adversarial Autoencoder-based Beamforming Design |
作者: | 陳弘運 HONG-YUN CHEN |
指導教授: | 周承復 Cheng-Fu Chou |
關鍵字: | 第六代 (6G),加性注意力,對抗自動編碼器,波束成形,毫米波,智慧反射面 (IRS), Sixth-generation (6G),additive attention,adversarial autoencoder,beamforming,millimeter-wave,Intelligent reflecting surface (IRS), |
出版年 : | 2023 |
學位: | 博士 |
摘要: | 6G 技術在速度、延遲和容量方面都超過了 5G,並引入了至關重要的智慧反射面(IRS)。在微控制器的管理下,這種高性價比的無源元件陣列可通過精確操縱傳入的無線電波來優化無線通訊,從而提高網路覆蓋、容量和能效。然而,現實世界中 IRS 的混合波束成形面臨著雜訊和干擾的困難。為了處理這個困難,我們提出了 AAE-AATT-波束成形(Adversarial AutoEncoder with Additive ATTention Beamforming)。加性注意力是一種強大的機制,用於在時域中對輸入序列中元素全域依賴性。AAE 學習到潛在空間在捕捉頻域和空間域通道內波束成形的基本特徵扮演重要角色。在類比預編碼模組中,自動編碼器被用來優化潛空間,並忠實地重建潛空間以匹配原始輸入通道信號資料,從而顯著提高捕捉全域特徵的準確性。數位預編碼模組利用具有平移不變性等特徵的 2D-CNN 捕獲通道預編碼的基本頻率和空間特徵,同時最大限度地減少干擾。在類比波束成形模組中,使用了門控遞迴單元(GRU),其重定和更新門控制單元內的資訊流。這提高了特徵捕捉的準確性。數位波束成形模組採用 1D-CNN 技術,善於捕捉連續資料中的局部模式和依賴關係,因此適用於時間序列分析等任務。該模組能有效捕捉關鍵通道波束成形特徵,同時減少干擾。實驗數值顯示和之前的研究對照,MSE、可實現速率、泛化性和強健性都有大幅提高。 The 6G technology, surpassing 5G in speed, latency, and capacity, introduces the crucial intelligent reflecting surface (IRS). Managed by a microcontroller, this cost-effective array of passive elements optimizes wireless communication by precisely manipulating incoming radio waves, enhancing network coverage, capacity, and energy efficiency. However, real-world hybrid beamforming in the IRS faces challenges from noise and interference. To tackle this issue, we present AAE-AATT-Beamforming (Adversarial AutoEncoder with Additive ATTention Beamforming). Additive attention is a powerful mechanism for modeling global relationships among elements within an input sequence in the time domain. AAE learned latent space plays a pivotal role in capturing essential features for intra-channel beamforming in both frequency and spatial domains in frequency and spatial domains. In the analog precoding module, an autoencoder is utilized to optimize the latent space and faithfully reconstruct it to match the original input channel signal data that significantly enhancing the accuracy of capturing global domain features. The digital precoding module utilizes a 2D-CNN with features like translation invariance that enabling it to capture essential frequency and spatial features for channel precoding while minimizing interference. Within the analog beamforming module, a Gated Recurrent Unit (GRU) is used, featuring reset and update gates that control information flow within the cell. This enhances feature capture accuracy. The digital beamforming module employs a 1D-CNN, adept at capturing local patterns and dependencies in sequential data, making it suitable for tasks like time series analysis. This module effectively captures key channel beamforming features while reducing interference. Numerical results demonstrate substantial improvements in MSE, achievable rate, generalizability, and robustness compared to prior research. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91228 |
DOI: | 10.6342/NTU202304239 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf | 6.11 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。