近年来，随着无人机在诸多领域取得的应用，由其产生的潜在安全威胁也逐渐受到关注。因此，作为实现无人机监管的重要前提，对无人机，尤其是非合作无人机的检测与识别正成为研究的热点。在无人机检测识别方法中，基于无人机无线信号进行检测与识别具有显著的优势；其中，利用机器学习手段辅助传统信号处理过程的射频机器学习方法正展现出巨大的发展前景。 本文旨在使用射频机器学习技术，对无人机无线信号识别问题进行研究，从而为进一步的应用奠定基础。具体的工作内容主要包括：首先，本文搭建了无线信号采集系统并对多台无人机信号进行了数据采集，构建了相应的数据集。其次，针对无人机控制信号的特点，使用了包括时频图变换、跳频图案识别和跳频脉冲位置估计在内的预处理方法。然后，提出了一套基于射频机器学习方法的层次化无人机控制信号识别算法框架，不但可以实现无人机的类型分类，还能对同一类型的单个无人机进行识别：在类型识别阶段，本文训练了一个针对跳频信号时频图特点专门设计的双流卷积神经网络对来自不同类型无人机的信号实现了分类；在个体识别阶段，通过估计脉冲结构、计算分形维数降维和训练支持向量机分类器来对来自同一型号无人机的控制信号进行个体识别。实验证明，在包含七架大疆无人机真实信号数据集上，本文提出的预处理及层次化无人机控制信号识别方法能够对信号进行检测和识别，在类型识别和个体识别两个阶段都能达到了99%以上的分类准确率，且在低信噪比情况下也有良好的表现。
In recent years, with the application of Unmanned Aerial Vehicles (UAVs) in many fields, the potential security threats caused by UAVs have attracted more and more attention. Therefore, as an important prerequisite for UAV supervision, the detection and identification of UAVs, especially non-cooperative UAVs, is becoming a hot topic. Among the UAV detection and identification methods, those based on UAV wireless signals have significant advantages. Among them, the Radio Frequency Machine Learning (RFML) method, which uses machine learning to assist the traditional signal processing, is showing great development prospects. This paper aims to research the wireless signal recognition of UAVs using the RFML technology, so as to become the basis of further applications. The specific work contents mainly include: firstly, we establish a wireless signal acquisition system, collects the signals of multiple UAVs, and constructs the signal dataset. Secondly, according to the characteristics of UAV control signals, a preprocessing method including spectrogram transformation, frequency hopping pattern recognition and frequency hopping pulse position estimation is proposed. Then, a hierarchical UAV control signal recognition algorithm framework based on RFML method is proposed, which can not only classify the type of UAV, but also identify UAVs of the same type: in the stage of type classification, a specially designed two-stream convolutional neural network is trained to classify the signals from different types of UAVs; In the individual identification stage, the control signals from the same type of UAVs are identified by estimating the pulse structure, calculating the fractal dimension and training the support vector machine classifiers. Experiments show that the preprocessing method and hierarchical framework proposed in this paper can detect and recognize the signal, achieve a classification accuracy more than 99% in both stages, and also performs well even at low signal-noise ratio levels.
