基于时频分析和深度学习的复杂调制雷达信号脉内调制方式识别

Radar signals intra-pulse modulation recognition is widely used in electronic reconnaissance equipment such as electronic support measures, electronic intelligence, radar warning receiver and so on. Under the highly intensive radar signals environment of modern battlefield, the existing radar signal recognition methods cannot accomplish identification of complex modulation signals at low SNR, including diversified modulation signals, compound modulation signals, and overlapping signals in time and frequency domain.This project integrates some theories and methods such as time-frequency analysis, image processing and deep learning. According to the characteristics of radar signal ambiguity pattern, a reasonable time-frequency analysis kernel function is designed to suppress noise and cross terms in time and frequency image. Image segmentation methods and mathematical morphology based signal separation method are proposed to solve the problem of multi-signal separation， and the problem of overlapping multi-signal recognition is transformed into the problem of single-signal recognition. The method of radar signal intra-pulse modulation recognition based on time-frequency image and deep learning is proposed. This method avoids the complex processes of feature extraction and selection in traditional recognition methods, meanwhile this method extracts optimal features of radar signals adaptively to adapt to more kinds of complex modulated radar signals and enhance the generalization ability of recognition methods. The research results of this project have important significance and practical value for upgrading the capabilities of China's electronic warfare technology and obtain the information dominance of the future information warfare.

雷达信号脉内调制类型识别在电子支援、电子情报和雷达威胁告警系统等电子侦察设备中具有广泛的应用。在现代战场高度密集的雷达信号环境下，现有雷达信号识别方法无法完成低信噪比下复杂调制信号包括多样化调制类型、复合调制、时频域交叠信号的识别。本项目融合时频分析、图像处理和深度学习等理论和方法，针对雷达信号模糊域特点，设计合理的时频分析核函数，抑制信号时频图中的噪声和交叉项；提出采用时频图像分割和基于数学形态学的信号分离方法解决多信号分离问题，将交叠多信号识别转化为单信号识别问题；提出基于雷达信号时频图和深度学习的脉内调制类型识别方法，避免传统识别方法中复杂的特征提取和选择过程，完成雷达信号特征自适应最优提取，以适应更多种类的复杂调制雷达信号，提升识别方法的泛化能力。本项目的研究成果对提升我国电子战技术能力和水平，抢占未来信息战的制信息权具有重要的意义和实用价值。

雷达信号脉内调制类型识别在电子支援、电子情报和雷达威胁告警系统等电子侦察设备中具有广泛的应用。在现代战场高度密集的雷达信号环境下，现有雷达信号识别方法无法完成低信噪比下复杂调制信号包括多样化调制类型、复合调制、时频域交叠信号的识别。本项目提出了基于雷达信号时频图和深度学习的脉内调制类型识别方法，避免了传统识别方法中复杂的特征提取和选择过程，完成了雷达信号特征自适应最优提取，以适应更多种类的复杂调制雷达信号，提升识别方法的泛化能力。针对雷达信号特点，提出了基于多核函数的Cohen类时频分布和曝光融合算法的方法，获取多分量雷达信号的时频图像，抑制噪声和交叉项，且最大限度保留雷达信号的时域和频域特征；提出了基于卷积去噪自编码器的雷达信号时频图像去噪方法，可以有效地去除图像噪声，保留输入的时频图像原有信号信息，提高识别算法的抗噪能力，提升低信噪比下的雷达信号脉内调制分类的成功概率；提出了基于Inception模块和卷积去噪自编码器的单分量复杂调制雷达信号识别算法，在信噪比-9dB时，对十二种复杂调制雷达信号识别正确率达到90%以上，并且该算法具有很强的抗噪性能和泛化性能；在多分量雷达信号识别方面，提出了基于强化学习的多分量雷达信号分类识别算法，实现了8类随机交叠的典型雷达信号的脉内调制方式识别，在信噪比为-6dB时，算法的总体成功识别率达到了94.6%，同时该算法对单分量雷达信号的识别问题也具备适应性，在信噪比为-6dB时，算法的总体成功识别率达到了94.9%，解决了深度学习网络对时频域交叠的多分量雷达信号有效识别问题。.在项目执行期间，发表了高水平学术论文6篇，其中SCI收录6篇；申请国家发明专利8项；获得黑龙江省科技进步奖二等奖1项；培养了博士研究生2名，硕士研究生4名。