基于迁移学习的Vision Transformer网络对伽马射线暂现源分类  

Classification of gamma-ray transients using vision transformer network based on transfer learning

作　　者：张燕婷 马想[1] 黄跃[1] 刘佳聪 熊少林[1] 张鹏[1,3] 赵小芸[1] ZHANG YanTing;MA Xiang;HUANG Yue;LIU JiaCong;XIONG ShaoLin;ZHANG Peng;ZHAO XiaoYun(Institute of High Energy Physics,Chinese Academy of Sciences,Beijing 100049,China;University of Chinese Academy of Sciences,Beijing 101408,China;College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China)

机构地区：[1]中国科学院高能物理研究所,北京100049 [2]中国科学院大学,北京101408 [3]同济大学电子与信息工程学院,上海201804

出　　处：《中国科学：物理学、力学、天文学》2024年第8期100-110,共11页Scientia Sinica Physica,Mechanica & Astronomica

基　　金：国家自然科学基金(编号:12273042);国家重点研发计划(编号:2022YFF0711404);中国科学院空间科学战略性先导科技专项(编号:XDA15360000)资助项目。

摘　　要：伽马射线暂现源包括伽马射线暴(GRB)、软伽马射线重复暴(SGR)等极端天体爆发现象,主要由空间伽马射线监测器(例如GECAM卫星)开展探测研究,是天文学重要的研究前沿.另外,太阳活动和复杂的空间环境也会在空间伽马射线监测器产生各类爆发事件(统称为触发).在诸多触发事件中对GRB和SGR进行快速分类识别是开展后续物理研究的重要前提.传统的分类研究常使用贝叶斯方法,利用多个特征信息建模计算各个类别概率,但其需要大量的先验知识与合理的模型设想,且不同卫星间的传统分类模型面临迁移难、适用窄等问题.为高效快速地对触发进行分类,本文提出利用Vision Transformer网络,基于迁移学习和单流多模态的思想,使用较少的计算资源和数据对触发进行分类.本文将该方法应用于GECAM触发数据,实验结果表明,模型在测试集中准确率达到了89%,且模型能够理解类别间的区分与相似性.本文结果表明,这种新的方法对于伽马射线暂现源以及其他天文现象的分类具有良好应用前景.Gamma-ray transient sources include extreme astrophysical phenomena such as gamma-ray bursts(GRBs)and soft gamma-ray repeaters(SGRs).These phenomena are mainly studied using space gamma-ray detectors(such as the GECAM satellite)and represent an important frontier in astronomy research.Additionally,solar activity and complex space environments can generate different types of outbursts(collectively referred to as triggers)in space gamma-ray detectors.Among various trigger events,the rapid classification and identification of GRBs and SGRs are important prerequisites for subsequent physical research.Traditional classification studies often use Bayesian methods to model and calculate the probabilities of different categories using multiple feature information.However,these methods require considerable prior knowledge and reasonable model assumptions,and traditional classification models between different satellites face problems such as difficult migration and narrow applicability.To efficiently and rapidly classify triggers,this paper proposes using the vision transformer network,based on transfer learning and the single-stream multimodality concept,to classify triggers using fewer computational resources and data.We applied this method to GECAM trigger data.Experimental results show that the model achieved an accuracy of 89%in the test set and understood the similarity and distinction between categories.These results indicate that the proposed method has good application prospects for classifying gamma-ray transient sources and other astronomical phenomena.

关 键 词：伽马射线暂现源 迁移学习 触发分类 

分 类 号：P172.3[天文地球—天文学] TP18[自动化与计算机技术—控制理论与控制工程]