基于深度学习的信号调制识别研究

DEEP LEARNING BASED SIGNAL MODULATION RECOGNITION

石彦坤

11849056

硕士

信息与通信工程

贡毅

2020-05-29

2020-05-29

哈尔滨工业大学

深圳

调制识别（Modulation Recognition，MR）是在通信系统中介于信号检测和信号调制的一种技术，其主要目的是通过接收到的信号来分析信号调制方式，以便对信号进行后续的研究和处理。调制识别技术在电子对抗、无线电管理等领域都有着极其重要的作用，特别是在非合作通信方面。调制识别技术经过几十年的快速发展在理论方面已经十分成熟了，但随着无线通信环境的日益复杂，在实际的工程应用上仍有许多需要改进的地方。近几年来随着深度学习在多个领域的突出表现，研究学者开始关注深度学习与无线通信相关技术的结合与应用。基于深度学习的信号调制识别算法能够从不同类型的调制信号数据中自动学习到更深层次的特征表达，故本论文主要基于深度学习的方法来提高对信号调制识别的准确率。本论文首先对深度学习中的网络结构、优化算法以及训练模式等方法进行了学习研究。然后给出了11种调制信号（包括3种模拟调制信号和8种数字调制信号）的数学模型，并对这11种调制信号的复数信号进行分析，得到基于高阶累积量和时频分析的两种信号表征。此外，针对低信噪比环境下调制识别的准确率较低的问题，本论文结合时频分析的表征方法与基于条件生成对抗网络的降噪方法实现了对调制信号图像上噪声的减弱。最后，本论文对提出的图像降噪方法在公开数据集上与现有存在的算法进行对比，验证了信号表征方法可以有效地表达不同调制信号的特点，以及提出的降噪模型可以大幅度地提高低信噪比环境下的识别准确率。

Modulation recognition (MR) is a technology between signal detection and signal modulation in the communication system. The main purpose is to analyze the signal modulation mode through the received signal in order to carry out subsequent research and processing on the signal. Modulation recognition technology is extremely important in the fields of electronic countermeasures and radio management, especially in non-cooperative communication. After decades of rapid development, modulation recognition technology has been very mature in theory, but with the increasing complexity of the wireless communication environment, there are still many areas for improvement in practical engineering applications.In recent years, with the outstanding performance of deep learning in many fields, researchers have begun to pay attention to the combination and application of deep learning and wireless communication related technologies. The modulation signal recognition algorithm based on deep learning can automatically learn deeper feature expressions from different types of modulated signal data, so this thesis is mainly based on deep learning methods to improve the accuracy of signal modulation recognition.In this thesis, we first studied the network structure, optimization algorithm and training methods in deep learning. Then, the mathematical models of 11 kinds of modulation signals (including 3 kinds of analog modulation signals and 8 kinds of digital modulation signals) are given, and the complex signals of these 11 kinds of modulation signals are analyzed to obtain two signal representation based on high-order cumulants and time-frequency analysis. In addition, for the problem of low accuracy of modulation recognition in low signal-to-noise ratio environment, this paper combines the time-frequency analysis with the noise reduction method based on conditional generation adversarial network to reduce the noise on the modulated signal images. Finally, this paper compares the proposed image noise reduction method with the existing algorithms on public data sets, verifies that the signal representation method can effectively express the characteristics of different modulated signals, and the proposed noise reduction model can greatly improve the recognition accuracy under low signal-to noise radio environment.

调制识别 信号表征 生成对抗网络 卷积神经网络

modulation recognition signal representation generative adversarial network convolutional neural network

中文

联合培养

南方科技大学

石彦坤. 基于深度学习的信号调制识别研究[D]. 深圳. 哈尔滨工业大学,2020.