面向空战对抗行为意图分析的小样本学习方法  

A Few-shot Learning Method for Intent Analysis of Air Combat Confrontation Behaviors

作　　者：潘明 郑景嵩 李金亮 方龙 杨阳[2] 赵世杰[2] PAN Ming;ZHENG Jingsong;LI Jinliang;FANG Long;YANG Yang;ZHAO Shijie(Southwest China Research Institute of Electronic Equipment,Chengdu 610036,China;School of Automation,Northwestern Polytechnic University,Xi′an 710129,China)

机构地区：[1]中国西南电子设备研究所,成都610036 [2]西北工业大学自动化学院,西安710129

出　　处：《哈尔滨理工大学学报》2024年第4期50-58,共9页Journal of Harbin University of Science and Technology

基　　金：电磁空间实验室基金;国家自然科学基金(42271315);陕西省自然科学基础研究计划面上项目(2023JCYB559).

摘　　要：针对目前空战对抗中空战目标的行为意图识别存在着数据来源多、数据模态多、数据的维度高冗余大、样本量小和不均衡以及训练所需的大量标注数据获取困难等问题,构建了一种基于深度双向门控循环单元(deep bidirectional gated recurrent unit,DBGRU)的空战目标行为意图识别模型,通过在双向门控循环单元(bidirectional gated recurrent unit,BiGRU)中融合注意力机制来提升模型的特征学习能力,自适应地分配不同空战特征信息的权重。并以DBGRU为骨干网络,提出了一种基于数据扩充的小样本对比学习算法,利用基于Wasserstein距离的生成对抗网络(wasserstein generative adversarial network,WGAN)扩充原始数据,并利用对比学习框架挖掘多模态数据中的丰富的模式信息弥补小样本数据规模不足的缺陷,从而准确预测空战目标行为意图。实验仿真结果表明,基于数据扩充的小样本对比学习算法预测小样本空战目标行为意图的准确率为91.13%。Aiming at the problems of multiple data sources,multiple data modes,high data dimensions,large redundancy,small and unbalanced sample size,and difficulty in obtaining a large number of labeled data required for training,a deep bidirectional gated recurrent unit(DBGRU)electromagnetic behavior intent recognition model is constructed.By integrating the attention mechanism in the Bidirectional Gated Recurrent Unit(BiGRU),the feature learning ability of the model is improved,and adaptively assign the weight of different air combat feature information.With DBGRU as the backbone network,a few-shot contrastive learning algorithm based on data augmentation is proposed,which uses the Wasserstein Generative Adversarial Network(WGAN)based on Wasserstein distance to enrich the original data,and uses the contrastive learning framework to mine the rich pattern information in the multimodal data to make up for the lack of few-shot data,so as to accurately predict the behavior intention of electromagnetic targets.The experimental simulation results show that the accuracy of the few-shot contrastive learning algorithm based on data augmentation in predicting the behavior intention of few-shot air combat targets is 91.13%.

关 键 词：意图识别 注意力机制 门控循环单元 生成对抗网络 对比学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]