针对已有Cohen类时频分布等方法时频聚焦能力不足、在低信噪比(signal to noise ratio，SNR)情况下调制识别准确率低的问题，提出一种基于同步提取变换(synchro-extracting transform，SET)去噪的分组卷积神经网络调制识别方法。所提方法使用SET对雷达信号进行时频分析，以获得良好的时频聚焦性，提高时频分析的计算效率;通过Viterbi算法搜索估计时频系数矩阵中的瞬时频率轨迹，综合考虑信号能量强度分布与瞬时频率轨迹的平滑性，并对得到的瞬时频率轨迹进行中值滤波以去除脉冲噪声;保留瞬时频率轨迹邻域的时频系数，以达到时频图去噪的目的。最后，将去噪后的时频图送入具有残差连接的分组卷积神经网络进行特征提取与调制识别。实验结果表明，当SNR为-12 dB时，去噪后的SET时频图时频聚焦性好，调制识别准确率比未去噪的识别准确率提高了 13。69％，证明所提出的雷达信号调制识别方法在低SNR条件下对多种复杂调制类型的信号具有良好的识别性能。
Radar signal modulation recognition method based on synchro-extracting transform denoising
Due to the insufficient time-frequency focusing ability of the existing Cohen-class time-frequency distribution and low modulation recognition accuracy under low signal to noise ratio(SNR),a grouped convolutional neural network modulation recognition method based on synchronous extracting transform(SET)denoising is proposed.Firstly,SET is used for time-frequency analysis of the radar signals,providing better time-frequency focusing and computational efficiency of time-frequency analysis.Then,the Viterbi algorithm is utilized to search and estimate the instaneous frequency trajectory in the time-frequency coefficient matrix,taking into account the distribution of signal energy intensity and the smoothness of the instaneous frequency trajectory.At the same time,a median filter is applied to remove pulse noise from the obtained instaneous frequency trajectory,and the time-frequency coefficients in the vicinity of the instaneous frequency trajectory are retained to achieve time-frequency image denoising.Finally,the denoised time-frequency images are sent to a grouped convolution neural network with residual connections for feature extraction and modulation recognition.The experimental results demonstrate that,when the SNR is-12 dB,the denoised SET time-frequency images have good time-frequency focusing,and the modulation recognition accuracy is improved by 13.69％compared to the recognition accuracy without denoising.The proposed radar signal modulation recognition method exhibits excellent recognition performance for various complex modulation types of signals under low SNR conditions.
radar signal modulation recognitionsynchro-extracting transforminstaneous frequency estimationdenoising
NETL
NSTL
万方数据
