| 摘要: | 在目前的5G網路中，為了得到更高的資料吞吐量，常使用高頻段的頻譜來進行通訊，由於高頻段的訊號比低頻段的訊號有著更高的衰退率，進而導致5G有著低覆蓋率的問題。在高山、海洋和沙漠等無法建置基地台之區域就無法享受5G帶來的便利。在後5G的設計理念中，為了解決低覆蓋率問題，結合低軌道衛星與地面網路形成互補，實現全球覆蓋網路。然而，低軌道衛星有著高度移動性以及遠距離傳輸的特性，使得都卜勒效應(Doppler effect)和路徑損失(Path Loss) 嚴重加劇，容易造成通訊通道處於不穩定的狀態，使用者因而無法解析出正確的訊號。在5G中，為了解決通道衰弱的問題，提出許多自適應技術(Adaptive Transmission Technique)。自適應技術運作在傳送端，且必須藉由接收端執行通道估計(Channel Estimation)並回傳的通道狀態資訊 (Channel State Information (CSI)) 進行參數調整。然而，在衛星5G網路中，因長距離衛星鏈路所造成的傳輸延遲使得接收端收到過時的CSI而做出錯誤的自適應決策。所以預測CSI的演算法越來越受到重視。然而，大部分的預測演算法需要收集大量的環境參數才能進行預測。在快速變化的衛星通道中使用者需要頻繁地收集新的環境參數，造成使用者高計算負擔。因此，在本篇論文中，我們提出一個適用於低軌道衛星網路的輕量CSI預測方法。我們使用少量歷史CSI與循環神經網路 (Recurrent Neural Network) 來預測未來CSI，以降低使用傳統通道估計的頻率。;In the current 5G network, to obtain higher throughput, a high-band spectrum is often used for 5G communication. Since the signal in high-band has a higher declining rate than the signal in the low-band, 5G encounter low coverage problem. Users in areas such as mountains, oceans, and deserts where base stations cannot be built are not able to enjoy the convenience brought by 5G. In the specification of beyond 5G, to cope with the problem of low coverage, combining low orbit satellites (LEO) with the terrestrial network to achieve global coverage network. However, LEOs have high mobility and ability to long-range transmission characteristic, which makes the Doppler effect and path loss severely aggravated, which easily makes the communication channel to be in an unstable state, therefore, users cannot detect the accurate signal. In 5G, to solve the problem of the above mentioned, several Adaptive Transmission Technique are proposed. Adaptive transmission techniques usually operate on the transmitter side, and the adjustment of which is tuning by the Channel State Information (CSI), which is obtained through channel estimation and feedbacked from the receiver side. However, in the 5G satellite network, the obtained CSI might be already outdated before its actual usage due to the long-range satellite link. Therefore, the CSI prediction algorithm has gained more research interests. Most of the prediction methods need to collect a great number of environmental parameters to complete the predict process. In a time-variant satellite channel, the obtained estimates expire quickly with the change of propagation environments, which will cause high computational cost. In this paper, we propose a lightweight CSI prediction method for LEO communication. We use a small amount of historical CSI to train the Recurrent Neural Network (RNN) and predict the future CSI to reduce the use of traditional channel estimation. |
