面向MIMO系统在时变信道中的盲调制识别算法研究

The application of modulation recognition in MIMO systems are facing three main challenges, namely, high computation complexity, the dependence on channel state information and the poor adaptability in time-variant channels. The state-of-the-art likelihood base modulation recognition algorithms suffers significantly from high computational complexity. Although feature based method can reduce the computational complexity, it requires channel state information for channel equalization, which is impracticable in real world scenarios. In the meantime, there is no literature focusing on the research of modulation recognition in time-variant channels. Recently the application of machine learning algorithms in modulation recognition has received increasing amount of attention. According to recent studies, machine learning algorithms possess unique ability to uncover the hidden structural feature of modulated signals as compared to classic method. Thus, this project proposes to solve the problem of modulation recognition in MIMO time-variant channels via the combination of convolutional neural networks and transfer learning. This project is expected to significantly advance the research on information security and intelligent communication systems.

调制识别在MIMO系统中的应用面临着计算复杂度、信道信息依赖以及时变信道适应性三方面的技术挑战。经典算法中基于似然度的调制识别算法在MIMO系统中存在计算复杂度极高的问题。基于专家特征的识别算法有较低的计算复杂度，但需要精确的信道信息实现信道均衡，缺乏实际应用场景中的可行性。而对时变信道中调制识别问题领域内尚无具有实效解决方案。与此同时，近年来机器学习在调制识别方面的应用受到了越来越多的关注。同时众多相关研究显示机器学习算法在挖掘调制信号隐藏结构特征方面体现出了经典算法所不具备的独特优势。因此本项目拟采用卷积神经网络与迁移学习相结合的方法来解决调制识别研究在MIMO系统中时变信道下的核心问题。本项目的研究将对我国信息安全以及智能通信系统的发展起到重要的推动依赖

调制识别在军事电子对抗与民事智能通信系统中都有重要应用。在本项目中，我们探索了使用深度学习解决调制识别问题的各种可能方案。首先，我们设计了针对调制识别问题的具有特殊网络层功能定义的网络结构。其次，项目针对不同的信道条件提出了不同的卷积核初始化方法。最后，项目引入半监督学习与迁移学习尝试解决时变信道的问题。基于仿真实验结果，可以发现项目提出的网络结构能够成功解决调制识别的问题。且其中各层结构如预期实现不同信号分布特征的提取。算法在高斯白噪声信道、衰落信道中体现出了较强的鲁棒性。 半监督学习机制的引入大大减少了算法在训练过程中对大量已知信号的需求。迁移学习机制则在信道条件变化的情况下大幅减少了需要进行再训练的参数数量，同时训练获得的网络再调制识别性能方面并没有收到太大影响。结合上述研究内容，本项目成功提出了高度整合的基于卷积神经网络的调制识别方法。