基于SSTransformer的恒星亚型光谱分类方法研究  

Research on Spectral Classification of Stellar Subtypes Based on SSTransformer

作　　者：范雅雯 刘艳萍[1] 邱波 姜霞[1] 王林倩 王坤[1] FAN Ya-wen;LIU Yan-ping;QIU Bo;JIANG Xia;WANG Lin-qian;WANG Kun(Hebei University of Technology,Tianjin 300400,China)

机构地区：[1]河北工业大学,天津300400

出　　处：《光谱学与光谱分析》2023年第8期2523-2528,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金天文联合基金项目(U1931134);河北省自然科学基金面上项目(A2020202001)资助。

摘　　要：恒星的分类问题一直是天文研究的一大热点,恒星的亚型分类对探究恒星演化、稀有天体识别等具有重大意义。针对LAMOST光谱亚型分类问题设计了SSTransformer(stellar spectrum transformer)分类模型,该模型主要由三部分组成,包括输入模块、嵌入模块、SST编码模块。在输入模块中,将光谱数据进行分块处理,这些块经过线性投射层被映射为向量。在嵌入模块中,为了提取有用的数据特征,将线性投射层的输出加入一个可学习的类别嵌入块,为了保留位置信息,再加入位置嵌入块,之后将这些数据特征向量送入SST编码模块。最后在SST编码模块中,对数据特征进行提取处理,并利用多层感知器结合新特征对恒星光谱进行分类。采用的A、F、G、K、M型恒星光谱数据均来自LAMOST DR8中的一维低分辨率光谱,35256条一维光谱数据用于SSTransformer模型的训练,8815条一维光谱数据用作模型的测试。为了加快模型的收敛速度,对数据采用Z-Score归一化处理。由于是多分类问题,实验采用了准确率、精确率、召回率、F1-Score、Kappa系数五个评价指标。实验结果证明,利用SSTransformer模型可实现对一维恒星光谱数据有效的筛选分类,分类准确率达到98.36%,比支持向量机(support vector machine,SVM)算法、极端梯度提升(eXtreme Gradient Boosting,XGBoost)算法,以及卷积神经网络(convolutional neural networks,CNN)的分类准确率更高。The classification of stars has always been a hot topic in recent astronomical research.The classification of stellar subtypes is significant for exploring stellar evolution and identifying rare celestial bodies.This paper designs the SSTransformer(Stellar Spectrum Transformer)classification model for the LAMOST spectral subtype classification problem.The model mainly comprises three parts,including the input module,the embedding module,and the SST encoding module.In the input module,the spectral data is processed into blocks,which are mapped to vectors through a linear projection layer.In the embedding module,in order to extract useful data features,the output of the linear projection layer is added to a learnable category embedding block.In order to preserve the position information,a position embedding block is added,and then these data feature vectors are sent to the SST encoding module.Finally,the data features are extracted in the SST coding module,and the stellar spectrum is classified using the multilayer perceptron combined with the new features.In this paper,the spectral data of type A,F,G,K,and M starsis all from the one-dimensional low-resolution spectra in LAMOST DR8,35256 pieces of one-dimensional spectral data are used for training the SSTransformer model,and 8815 pieces of one-dimensional spectral data are used for testing the SSTransformer model.In order to speed up the convergence of the model,Z-Score normalization is used for the data.Because this is a multi-classification problem,the experiment adopts five evaluation indicators:accuracy rate,precision rate,recall rate,F1-Score,and Kappa coefficient.The experimental results show that the SSTransformer model can effectively screen and classify one-dimensional stellar spectral data,and the classification accuracy reaches 98.36%,which is higher than the support vector machine(SVM)algorithm,eXtreme Gradient Boosting(XGBoost)algorithm,and convolutional neural networks(CNN).

关 键 词：恒星光谱 自动分类 SSTransformer模型 归一化 

分 类 号：P157.2[天文地球—天文学]