【作者】 王渊；

【导师】 李渝舟；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2020， 硕士

【摘要】 自动调制识别(Automatic Modulation Classification,AMC)是指在没有或者缺乏调制参数以及链路特性等先验信息的条件下用于评估未知信号调制类型的一种信号处理技术,在电子侦察、通信对抗、干扰识别、频谱监测等军事和民用领域有着广泛的应用。近年来,随着无线通信技术的迅猛发展,信号调制方式日益增加,信道传输环境也日趋复杂,这极大的增加了调制识别的难度。为此,如何设计一种高效的调制识别方法成为无线通信领域亟待解决的难题。本文致力于对多种调制类型、强干扰信道条件下的无线通信信号进行调制识别研究,通过将基于深度学习的AMC方法和基于特征提取的模式识别方法相结合以提升整体算法的调制识别性能。具体地,本文的研究内容和成果如下:1、本文通过挖掘调制信号的时空特性,设计了一种基于卷积长短期记忆(Convolutional Long Short-Term Memory,ConvLSTM)网络的调制识别模型。考虑到网络结构设计的重要性,本文通过多组仿真对比实验探讨了ConvLSTM层数、层次连接关系、神经元数目组合等多种网络参量对于调制识别效果的影响。基于此,我们得到了一个合适的网络结构,并设计了一种基于ConvLSTM的AMC算法。此外,为了进一步验证所构建网络模型的调制识别性能,本文基于广泛使用的RML2016.10a调制数据集将ConvLSTM网络与卷积神经网络(Convolutional Neural Network,CNN)、卷积长短期记忆深度神经网络(Convolutional,Long Short-Term Memory,Fully Connected Deep Neural Networks,CLDNN)的调制识别性能进行了对比,结果表明,基于ConvLSTM的网络表现出较好的识别效果。2、考虑到基于ConvLSTM的调制识别网络对于十六进制正交幅度调制(16Quadrature Amplitude Modulation,16QAM)与六十四进制正交幅度调制(64 Quadrature Amplitude Modulation,64QAM)、双边带调制(Amplitude Modulation-Double Side Band,AM-DSB)与宽带调频(Wide Band Frequency,WBFM)两组信号难以实现有效的区分,本文选用基于特征提取的模式识别方法来解决该问题。通过对[16QAM,64QAM]、[AM-DSB,WBFM]两组信号进行统计特性分析,依次提取了信号的信息熵、功率谱密度峰值、瞬时幅值标准偏差三种高区分度的特征以构建特征空间,并采用随机森林分类器对所构造的特征空间进行调制识别。结果表明,采用该方法可大大提升算法对于[16QAM,64QAM]、[AM-DSB,WBFM]两组信号的识别效果。3、为了实现真正意义上的自动调制识别,本文将深度学习与特征提取两种方法相结合,提出了一种基于级联网络构架的无线通信信号智能调制识别算法。具体地,首先将RML2016.10a调制数据集中的16QAM和64QAM信号合并为“QAM”信号,将AM-DSB与WBFM信号合并为“AW”信号;然后使用基于ConvLSTM的网络构建第一级调制识别模型来识别RML2016.10a数据集中的9种信号;接下来对ConvLSTM网络预测得到的AW和QAM信号提取信息熵、功率谱密度峰值、瞬时幅值标准偏差三种特征组成三维特征向量;最后利用随机森林分类器构建第二级调制识别模型来识别[16QAM,64QAM]、[AM-DSB,WBFM]两组信号。通过这种方式可以实现对数据集中11种调制信号的有效识别。实验结果表明,相比基于CNN、CLDNN和ConvLSTM 11分类的网络,本文提出算法的调制识别性能有了很大程度的提升,就平均识别率和最高识别率而言,相比CNN网络分别提升了11.53%和16.71%,相比CLDNN网络分别提升了10.63%和14.01%,相比ConvLSTM 11分类网络分别提升了5.53%和8.81%。更多还原

【Abstract】 Automatic modulation classification(AMC)is a kind of signal processing technique used to determine the modulation type of wireless communication signals in the absence of prior information such as modulation parameters and link characteristics.Nowadays,AMC technology is widely used in military and civil fields such as electronic reconnaissance,communication countermeasure,interference identification,and spectrum monitoring.For the past few years,with the fast development of communication technology,the modulation types of communication signals are increasing,and the channel transmission environment is becoming more and more complex,which greatly increase the difficulty of modulation classification.Therefore,how to design an efficient modulation classification method has become a hot topic to scholars.We are committed to the research of modulation classification of wireless communication signals under the condition of multiple modulation types and strong interference channels.By combining the deep learning and feature extraction methods,we propose an intelligent modulation classification algorithm for wireless communication signals,and achieve satisfactory classification performance.The research contents and achievements are as follows:1.We design a modulation classification network based on the convolutional long short-term memory(ConvLSTM)to extract the underlying temporal and spatial characteristics of modulation signals.Taking into account the importance of network structure design,we thoroughly explore the impacts of the ConvLSTM layer number,network structure,and neuron number on the modulation classification performance through multiple sets of comparative experiments.Based on this,we obtain a suitable network structure and then develop a ConvLSTM-based automatic modulation classifier.To further verify the classification performance of the constructed network,we compare the modulation classification performance of the ConvLSTM-based network with two classic networks based on convolutional neural network(CNN)and convolutional,long short-term memory,fully connected deep neural networks(CLDNN)by utilizing the widely-used RML2016.10 a modulation dataset.The results show that our proposed classifier exhibits better modulation classification performance.2.Considering that the ConvLSTM-based modulation classification network is difficult to effectively distinguish the 16 quadrature amplitude modulation(16QAM)and 64 quadrature amplitude modulation(64QAM),the amplitude modulation-double side band(AM-DSB)and wide band frequency(WBFM)two sets of modulation signals,we adopt the pattern recognition method based on feature extraction to solve this problem.By analyzing the statistical characteristics of [16QAM,64QAM] and [AM-DSB,WBFM] two sets of signals,three kinds of high distinguishing features of information entropy,peak value of power spectral density,and standard deviation of instantaneous amplitude are extracted to construct a three-dimensional feature space,and random forest classifier is used to perform classification operations on the constructed feature space.The experimental results indicate that this method can greatly improve the modulation recognition performance of [16QAM,64QAM] and [AM-DSB,WBFM] signals.3.To realize the automatic modulation classification in real sense,we propose a cascade network based intelligent modulation classification algorithm for wireless communication signals by combining the depth learning and feature extraction methods.Specifically,we first merge the 16 QAM and 64 QAM signals in the modulation dataset into “QAM” signals,and the AM-DSB and WBFM signals are merged into “AW” signals.We then utilize the ConvLSTM-based network mentioned above to build a first-level modulation classification model to identify nine kinds of modulation signals.Next,we extract the three features of information entropy,peak value of power spectral density,and standard deviation of instantaneous amplitude from AW and QAM signals predicted by the aforementioned ConvLSTM network to form a three-dimensional feature vector.Finally,the random forest classifier is used to construct the second-level modulation classification model to identify two groups of signals,i.e.,[16QAM,64QAM],[AM-DSB,WBFM].In this way,11 kinds of modulation signals in the dataset can be identified effectively.The experimental results show that,compared with the networks based on CNN,CNDNN,and ConvLSTM,the modulation classification performance of our proposed algorithm has been greatly improved.In terms of the average classification accuracy and the highest classification accuracy,our proposed algorithm is 11.53% and 16.71% higher than the CNN-based network,10.63% and 14.01% higher than the CLDNN-based network,and 5.53% and 8.81% higher than the ConvLSTM-based network.更多还原