【作者】 刘昕宇；

【导师】 孙晓东；

【作者基本信息】 吉林大学 ， 控制工程（专业学位）， 2022， 硕士

【摘要】 通信信号调制识别在军事领域及民用领域均具有重要的研究意义。在缺乏先验信息条件下,准确识别接收信号的调制方式对于接收端后续操作和处理十分关键。传统调制识别方法需要人为设计、筛选和提取特征,所提取的特征泛化能力较差,在低信噪比环境中,极易发生特征的失效,导致识别准确率大幅下降,甚至无法实现有效的识别。基于此,本文将信号处理与深度学习相结合,有效提升低信噪比下通信信号的调制识别效果。首先,本文对通信信号调制原理、信号时频表征方法以及深度学习相关理论进行了研究,详细介绍了非线性时频分析理论与卷积神经网络(CNN)的网络结构、常用优化算法和激活函数,为后续基于信号时频表征与卷积神经网络相结合的调制识别算法奠定理论基础。然后,设计了五种CNN结构,采用平滑伪Wigner时频分布生成时频图数据集用于网络的训练和测试,通过实验对比确定最佳网络——特征通道串联网络,为后续章节中网络结构的选择奠定基础。接着,针对低信噪比下,平滑伪Wigner分布时频图中噪声对有用信号特征干扰严重的问题,采用高阶时频分析方法——Wigner四阶矩谱作为信号的时频表征方式,通过切片操作实现降维,采用模糊域核函数滤波方法去除时频面上的交叉项,并分析该高阶分布的时频聚集性;通过时频图的仿真对比,体现低信噪比下高阶时频分布对有用信号特征具有更好的凸显能力;为了进一步提升网络识别性能,为网络添加ECA注意力机制;为了降低参数含量及计算复杂度,实现了轻量化改进。最后,同样针对平滑伪Wigner分布时频图受噪声干扰严重的问题,设计了基于残差学习的时频图去噪网络以及基于轻量化残差注意力UNet的时频图去噪网络。分别使用两种时频图去噪网络对带噪时频图数据集去噪,生成两种去噪时频图数据集,用于训练特征通道串联-ECA网络,验证该算法对识别效果的提升作用,并对比两种去噪网络的去噪性能。实验结果表明:在保持识别网络结构不变的情况下,本文设计的两种时频图去噪网络能在有效去除噪声分量的情况下,保留信号特征,提升低信噪比下的识别效果。综上,本文有效结合信号处理和深度学习方法,从改善信号时频图和设计、优化CNN网络结构两方面出发,提升低信噪下的识别效果。实验结果表明本文方法能显著提升低信噪比下的识别准确率。更多还原

【Abstract】 Modulation and identification of communication signals has important research significance in both military and civilian fields.In the absence of prior information,it is critical to accurately identify the modulation mode of the received signal for subsequent operations and processing at the receiving end.The traditional modulation recognition method requires artificial design,screening and extraction of features,and the extracted features have poor generalization ability.identify.Based on this,this paper combines signal processing with deep learning to effectively improve the modulation recognition effect of communication signals under low signal-to-noise ratio.Firstly,this paper studies the principle of communication signal modulation,signal time-frequency characterization method and related theories of deep learning,and introduces in detail the nonlinear time-frequency analysis theory and the network structure of convolutional neural network(CNN),commonly used optimization algorithms and activation functions,which lays a theoretical foundation for the subsequent modulation recognition algorithm based on the combination of signal time-frequency representation and convolutional neural network.Then,five CNN structures are designed,and the time-frequency map data set is generated by the smooth pseudo-Wigner time-frequency distribution for network training and testing,and the optimal network—the feature channel series network is determined through experimental comparison,which is the network structure in the subsequent chapters.The choice lays the foundation.Next,in view of the problem that the noise in the time-frequency diagram of the smooth pseudo-Wigner distribution seriously interferes with the characteristics of the useful signal under the low signal-to-noise ratio,a high-order time-frequency analysis method—Wigner fourth-order moment spectrum is used as the time-frequency representation method of the signal.The slicing operation realizes dimensionality reduction.The fuzzy domain kernel function filtering method is used to remove the cross terms on the time-frequency surface,and the time-frequency aggregation of the high-order distribution is analyzed.The time-frequency distribution has a better ability to highlight useful signal features;in order to further improve the network recognition performance,an ECA attention mechanism is added to the network;in order to reduce the parameter content and computational complexity,lightweight improvements are achieved.Finally,also for the problem that the time-frequency graph of smooth pseudo-Wigner distribution is seriously disturbed by noise,a time-frequency graph denoising network based on residual learning and a time-frequency graph denoising network based on lightweight residual attention UNet are designed.Two kinds of time-frequency map denoising networks are used to denoise the noisy time-frequency map data set,and two kinds of denoising time-frequency map data sets are generated,which are used to train the feature channel series-ECA network to verify the improvement of the algorithm on the recognition effect.and compare the denoising performance of the two denoising networks.The experimental results show that while keeping the structure of the recognition network unchanged,the two time-frequency graph denoising networks designed in this paper can effectively remove the noise components,retain the signal features,and improve the recognition effect under low signal-to-noise ratio.In summary,this paper effectively combines signal processing and deep learning methods to improve the recognition effect under low signal noise from two aspects:improving the signal time-frequency diagram and design,and optimizing the CNN network structure.The experimental results show that the proposed method can significantly improve the recognition accuracy under low signal-to-noise ratio.更多还原