本論文是將深度學習(DL)架構結合毫米波之多輸入多輸出的混合預編碼與合併器設計。在傳統的數位預編碼架構中，每根天線都會對應一個專用射頻鏈，但這在多輸入多輸出的系統中，天線數量將會非常龐大，所以混合預編碼正是改善傳統數位預編碼的問題，藉由巨量的訓練資料訓練，預測出類比預編碼的射頻鏈應有的形式，並與數位預編碼結合，降低所需的射頻鏈數量，達成降低硬體成本及高功率消耗的問題。 論文中，是採用半監督是學習的方式，預測出類比的射頻鏈。使用自編碼(Autoencoder)的神經網路設計，因自編碼的內神經有對稱性質，藉此找出相關的類比預編碼與結合器的射頻鏈，而我們可利用最小平方解獲得數位的預編碼與結合器。在不同的數據流情況下，從內神經網路結果產生的頻譜效率證明設計的推測是可執行的。 ;We apply deep learning (DL) structure to hybrid precoding and combining design in millimeter wave (mmWave) communication and multiple-input multiple-output (MIMO) system. In the traditional digital precoding structure, each antenna corresponds to a dedicated RF chain, the amount of antenna will be very large. Hybrid precoding and combining will improve traditional full digital precoding problem. By huge amount of training data, predict the analog precoding and combining RF chain, this can reduce the cost and power consumption. We try to use semi-supervised learning, predict the analog RF chain. Using Autoencoder neural network design because Autoencoder neurons have symmetric properties, to find the related analog precoding and combining RF chain, digital precoder and combiner can be calculated by least square solution. In the different cases of data streams, we calculate the spectrual efficiency by neural network, it can be proved that the design is feasible.
