基于时相特征卷积网络的雷达波形识别
Radar waveform recognition based on time and phase features convolutional network
崔天舒 1刘航 2范宇杰 2石亮 1张宏江1
作者信息
- 1. 中国航天科技创新研究院,北京 100080
- 2. 航天东方红卫星有限公司,北京 100089
- 折叠
摘要
低截获概率(LPI)雷达以其优越的反截获性能得到广泛应用,识别LPI雷达波形对电子侦察系统至关重要.然而,基于深度学习的LPI雷达波形识别方法存在网络参数量大、计算复杂度高等问题,严重影响了其在算力不足的场景下的使用.面向LPI雷达波形快速精确识别需求,提出了一种基于时相特征卷积网络的雷达波形识别模型.该方法与传统的时频变换方法不同,采用卷积神经网络直接从原始信号中提取相位和短时特征,具有轻量化和低复杂度的特点.通过包含 13个LPI雷达波形识别实验的验证表明,该方法即使在信噪比为-4 dB的情况下,也能实现90%以上的准确率;与基于Wigner-Ville变换和图像深度网络识别方法相比,该方法只需要12%的参数和0.14%的计算资源,就可达到相同精度,兼顾了准确率和处理速度,具有非常好的工程应用前景.
Abstract
Low probability of intercept(LPI)radars are widely employed due to their excellent anti-intercep-tion capabilities,and the identification of LPI radar waveforms is essential for electronic reconnaissance systems.However,the existing deep learning-based method for LPI radar waveform recognition faces challenges,such as large network parameters and high computational complexity.These issues significantly limit its applicability in re-source-constrained scenarios.Addressing the need for swift and accurate LPI radar waveform recognition,a novel radar waveform recognition model based on a time-phase feature convolutional network is introduced.Unlike tra-ditional time-frequency transformation methods,this approach utilizes a convolutional neural network to directly extract phase and short-time features from the raw signal.This results in a lightweight and low-complexity model.Validation through 13 LPI radar waveform recognition experiments demonstrates that the proposed method achieves over 90%accuracy even at a signal-to-noise ratio of-4 dB.Compared to recognition methods based on the Wigner-Ville transform and image deep networks,The presented method requires only 12%of the parame-ters and 0.14%of the computational resources to achieve equivalent accuracy which strikes a balance between ac-curacy and processing speed,exhibiting promising potential for practical engineering applications.
关键词
低拦截概率雷达/波形识别/深度学习/卷积神经网络/相位和短时特征Key words
low probability of intercept radar/waveform recognition/deep learning/convolutional neural network/phase and short-time feature引用本文复制引用
出版年
2024
NETL
NSTL
万方数据