基于Hilbert-Huang变换与对抗训练的特定辐射源识别

为有效解决特定辐射源的个体识别问题, 提出一种基于Hilbert-Huang变换与对抗训练相结合的方法。首先根据辐射源硬件差异, 建立辐射源信号的数学模型; 其次, 对信号进行Hilbert-Huang变换得到Hilbert谱; 然后, 在预处理过程中, 从信号所有的Hilbert谱时频点对应的能量值中, 确定最具区分度的一组能量值, 并记录其对应的时频点; 最后, 对每一类辐射源信号的Hilbert谱提取上述记录的时频点对应的能量值, 将其送入卷积神经网络进行训练与测试, 并通过对抗训练的方式提升网络的抗噪性能。识别准确率实验表明, 对比不进行对抗训练的方法以及不进行预处理与对抗训练的方法, 所提算法的识别率分别平均提升3.1%与5.45%。识别鲁棒性实验表明, 所提算法训练样本为100时即可达到较好识别效果, 同时随着辐射源个数增多优势更加明显。复杂度分析表明, 所提算法能有效降低神经网络在大量训练与识别过程产生的运算量。

Specific emitter identification based on Hilbert-Huang transform and adversarial training

In order to effectively solve the problem of specific emitter identification, a method based on the combination of Hilbert-Huang transform and adversarial training is proposed. Firstly, a mathematical model of the emitter signal according to the hardware difference of the emitter is established. Secondly, the Hilbert-Huang transform is performed on the signal to obtain the Hilbert spectrum. Then, in the preprocessing process, from the energy values corresponding to all Hilbert spectrum time-frequency points of the the signal, the most distinguishable set of energy values is determined, and the corresponding time-frequency points are recorded. Finally, the energy values corresponding to the time-frequency points recorded above for the Hilbert spectrum of each type of emitter signal is extracted. And then, it is sent to the convolutional neural network for training and testing, and the anti-noise performance of the network is improved by means of adversarial training. The identification accuracy experiment shows that comparing the method without adversarial training and the method without preprocessing and adversarial training, the identification accuracy of the proposed algorithm is increased by 3.1% and 5.5%, respectively. The identification robustness experiment shows that the proposed algorithm can achieve great identification performance when the training sample is 100, and the advantages become more obvious as the number of radiation sources emitter increases. The complexity analysis shows that the proposed algorithm can effectively reduce the amount of calculations generated by the neural network during a large number of training and recognition processes.