基于深度自编码器的辐射源个体开集识别

为实现对城市用频设备的精确管控，针对特定辐射源开集识别问题，构建了一套基于深度学习的辐射源个体开集识别处理流程，核心在于指纹特征有效区间筛选与基于深度自编码器的开集识别模型。一方面，通过使用Grad-CAM实现对深度网络激活可视化，筛选出信号对网络激活贡献较高的部分，在不损失过多指纹信息的情况下进行信号区间筛选；另一方面，建立基于半监督对抗自编码器的辐射源个体开集识别模型，实现对电磁环境中出现的未知辐射源个体的有效识别。实验表明此开集识别模型能够在不损失闭集识别率的条件下实现高精确度的开集识别。

Open set recognition of specific emitter identification based on deep auto-encoder

A processing process of open-set specific emitter identification is built in order to achieve accurate control of urban frequency equipment. The core lies in the effective interval filtering of fingerprint features and the open set recognition model based on the deep self-encoder. By visualizing deep network activation using Class Activation Mapping(Grad-CAM), the section of signal contributing more to neural network activation can be determined, and then interval filtering for the signal can be performed without losing too much fingerprint information. On the other hand, an open-set specific emitter identification model is established based on semi-supervised adversarial autoencoders, achieving effective monitoring and identification of unknown emitters that may occur in the spectrum. Experiments show that Grad-CAM can filter out the most advantageous part of the extracted signal fingerprint, and the proposed model can achieve high-precision open set recognition without degrading the closed set recognition rate.