| 引用本文: | 胡婉婉， 张劲东， 张 瑞， 王 娜. 基于轻量化网络的干扰信号智能识别算法[J]. 雷达科学与技术, 2023, 21(2): 133-142.[点击复制] |
| HU Wanwan, ZHANG Jindong, ZHANG Rui, WANG Na. Intelligent Identification Algorithm of Interference Signals Based on Lightweight Network[J]. Radar Science and Technology, 2023, 21(2): 133-142.[点击复制] |
|
| 摘要: |
| 针对干扰信号识别时卷积神经网络模型体积大、训练周期长、对硬件存储和计算要求高等问题，本文提出了一种基于改进轻量级网络模型SqueezeNet的干扰信号智能识别算法。该方法首先利用距离多普勒（Range?Doppler，RD）处理获得目标及干扰信号的RD图像，接着采用滑窗变换、归一化等技术对RD图像进行重塑；然后采用基于RD信息尺寸的隐层和通道剪枝措施对SqueezeNet网络进行改进，缩小了模型尺寸和存储空间；最后利用不同参数的测试样本拓展网络模型的泛化性能。仿真实验表明，在参数量减少到原网络1/30的情况下，改进的SqueezeNet网络对每种干扰的识别正确率可达到95%以上，且训练时间大大减少。 |
| 关键词: 干扰识别 距离多普勒图像 轻量化网络 参数压缩 |
| DOI：DOI:10.3969/j.issn.1672-2337.2023.02.003 |
| 分类号:TN974 |
| 基金项目:国家自然科学基金（No.62171220）； 上海航天科技创新基金（No.SAST2018?077） |
|
| Intelligent Identification Algorithm of Interference Signals Based on Lightweight Network |
|
HU Wanwan, ZHANG Jindong, ZHANG Rui, WANG Na
|
|
College of Electronic and Information Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China
|
| Abstract: |
| Aiming at the problems of large volume, long training time, and high requirement for hardware storage and calculation of convolutional neural network model in interference signal recognition, an intelligent identification algorithm of interference signal based on improved lightweight network model SqueezeNet is proposed in this paper. Firstly, range Doppler （RD） processing is used to obtain the RD images of targets and jamming signals under different parameters and the RD images are reconstructed by sliding window transform and normalization. Then, the hidden layer and channel pruning measures based on RD information size are used to improve the SqueezeNet network, reducing the model size and storage space. Finally, test samples with different parameters are used to expand the generalization performance of the network model. Simulation results show that when the parameters are reduced to 1/30 of the original network, the recognition accuracy of each kind of interferences can reach more than 95%, and the training time is also greatly reduced. |
| Key words: interference recognition range Doppler images lightweight network parameter compression |
