本篇論文中，提出了以卷積神經網路為基礎的數位調變分類器，卷積神經網路與其他深度學習結構相比，在圖像分類和語音辨識方面能給出更好的結果。吾人假設了單輸入單輸出系統及單輸入多輸出系統，傳送端發出QPSK、16-QAM、64-QAM三種不同的數位調變訊號，考慮通道衰減以及載波頻率偏移，接收端收到訊號後先經過所提出的預處理技術進行處理，包含通道的等化以及給定閾值以減少雜訊的影響，而當接收端的天線數為兩個以上時，會使用基於子空間的通道盲測方法對接收訊號進行處理，最後產生星座圖，作為卷積神經網路的訓練資料。 接著，吾人利用Zilinx Zynq-7000+AD9361軟體定義無線電平台，實現多天線傳輸的系統模型。訊號透過接收端的兩根天線接收後，會進行預處理及利用基於子空間方法的通道盲測進行還原，處理後的訊號會繪製於星座圖上，同樣作為卷積神經網路的輸入資料進行訓練。;In this paper, we propose a digital modulation classifier based on convolutional neural network. Compared with other deep learning architectures, convolutional neural network achieves better performance on the aspect of picture classification and speech recognition. In our simulation scenario, we assume SISO system and SIMO system, the transmitted signals are modulated by different types of digital modulation schemes, that is QPSK, 16-QAM and 64-QAM. We also consider channel fading and carrier frequency offset. At receiver, the proposed pre-processing techniques is employed, which realize channel equalization and retard the effect caused by noise. Also, when there are two or more antennas at receiver, we will utilize the subspace-based blind channel estimation to process received signals. Finally, we plot constellation diagrams according to the received signals after pre-processing. These constellation diagrams are the training data for the convolutional neural network model designed by us. Next, we use Zilinx Zynq-7000+AD9361 software-defined radio platform to implement real-time SIMO system. After collecting all signals received by the two antennas at receiver, we will make use of the proposed pre-processing techniques and subspace-based blind channel estimation to process and recover the received signals. Then, these complex sample points will be converted into constellation diagrams, and they are the training data for CNN.
