基于深度时频特征学习的雷达辐射源识别

针对雷达辐射源识别中拓展能力不足和识别率不高问题进行研究,提出一种基于深度时频特征学习的智能识别方法。基于降采样短时傅里叶变换高效提取具备较高辨识度和稳定性的浅层二维时频特征,利用信号局部频域维稀疏性完成降噪等预处理；设计用于深度特征学习与识别的卷积神经网络,并采用不同尺度卷积核组合扩展网络广度,强化特征表征能力；利用高信噪比条件下8种辐射源信号样本对网络进行训练调优,低信噪比样本测试验证算法和网络的有效性。仿真结果表明,该方式在-8 dB信噪比条件下能达到98.31%的整体平均识别率,具备较强的鲁棒性。

Radar emitter recognition based on the deep learning of time-frequency feature

Aiming at the problem of insufficient expansion ability and low recognition rate in radar emitter recognition, an intelligent recognition algorithm based on the deep learning of time-frequency feature was proposed. The shallow two-dimensional time-frequency features with high recognition and stability were quickly extracted by down sampling of short-time Fourier transform, and the noise reduction and other pre-processing were completed by using the sparseness of the local frequency-domain signal; a convolutional neural network for deep feature learning and recognition was designed, and the scale of the network was expanded by different scale convolution kernels to enhance the feature representation ability; the network was trained and tuned by using eight kinds of emitter signals under high SNR(signal-to-noise ratio) conditions, and the effectiveness of the algorithm and network was verified by a low SNR sample. The experimental results showed that the system achieves overall recognition rate of 98.31% at SNR of -8 dB, which verifies that the proposed algorithm has strong robustness.