基于卷积网络和对抗式学习的雷达信号识别算法

Radar Signal Recognition Algorithm Based on the Convolutional Neural Networks and Adversarial Learning

针对雷达信号识别中因新信号数据与训练数据的样本差异,导致卷积网络模型识别率下降的问题,提出一种基于卷积网络和对抗学习的雷达信号识别算法,提升模型对新样本的识别率。首先,依据时频变换理论,变换得到雷达信号时域-频域能量分布图像,以表征信号脉内调制特征;然后,结合深度可分离卷积的结构,构建卷积网络模型,并利用样本数据对模型进行预训练;最后,融合生成对抗网络的思想,网络模型采用对抗学习方式,根据输入新样本数据自适应调整、更新网络各层参数,提升模型对新样本特征提取能力。仿真结果表明,-4dB信噪比样本识别率由79.78%提升到90.67%,-2dB信噪比样本识别率由83.15%提升到91.00%,验证了算法的有效性。

In the process of radar signal recognition,there is a recognition rate decline which caused by the divergence between the new samples and the training samples.To improve the radar signal recognition rate,a algorithm based on convolutional networks and adversarial learning is proposed.First,the signal samples are transformed into time-frequency domain energy distribution figures by the time-frequency transformation.Then,the convolutional networks are built by the depthwise separable convolution units and trained by the time-frequency figures.At last,the parameters of the model are fine-tuned and updated adaptively by adversarial learning with the new samples,the abilities of feature extraction and the recognition rates of the model are increased with the low SNR new samples.The simulation results shows that the recognition rate increased from 79.78%to 90.67%at-4dB SNR,and from 83.15%to 91.00%at-2dB SNR,which verifies the effectiveness of the algorithm.