基于TransUnet的侵彻多层过载信号生成

Penetration Multilayer Overload Signal Generation Based on TransUnet

随着现代电子技术的蓬勃发展,人工智能在军事领域呈现快速上升的鳌头趋势,同时也面临着数据量严重不足的困境,针对目前侵彻多层过载信号数量少、类别不平衡等问题,结合深度学习的InfoGAN模型,提出一种基于TransUnet的侵彻多层过载信号生成方法。首先,以靶场实测过载信号作为训练模型的数据集,根据侵彻层数建立标签信息;其次,构建生成器与判别器,生成器借鉴TransUnet的思想,由Transformer Encoder和U-Net组成,用于学习过载数据集的特征映射,判别器则使用较为简单的注意力模型,以降低整体模型的复杂度;最后,利用生成对抗网络对生成器与判别器进行训练与优化,实现侵彻多层过载数据生成。实验结果表明,该方法能够根据不同层数信息生成不同速度的有效过载数据,可在一定程度上解决侵彻多层过载信号缺乏的问题。

With the booming development of modern electronic technology,artificial intelligence in the military field shows a hasty rising trend,but also faces the dilemma of a serious lack of data volume.To address the current problems of low number and unbalanced categories of penetration multilayer overload signals,we propose a TransUnet-based penetration multilayer overload signal generation method combined with the fruitful InfoGAN model of deep learning.Firstly,real overload signals at the range are used as the dataset of the training model,and the label information is established according to the number of penetration layers.Secondly,the generator and discriminator are constructed.The generator borrows the model structure of TransUnet,which consists of a Transformer Encoder and U-Net for learning the feature mapping of the overload dataset.At the same time,the discriminator applies a simpler attention model to reduce the overall model complexity.Finally,generators and discriminators are trained and optimized by using the generative adversarial network to generate multilayer penetration overload data.Experimental results show that the method can effectively overload data with different speeds according to the information of different layers.The problem of lack of penetration multilayer overload signals can be solved to a certain extent.