SATS：一种基于多重特征提取的恒星光谱分类算法

doi:10.3964/j.issn.1000-0593(2024)07-2029-08

摘要
参考文献
相关文章 (15)

全文: PDF (6937 KB)
输出: BibTeX | EndNote (RIS)

摘要：对恒星光谱进行深入研究，可以了解恒星的化学组成和物理特性。恒星光谱分类是恒星光谱研究领域的一个重要方向，随着海量恒星光谱数据的出现，人工分类手段就无法满足科研的需要。基于此，搭建了SATS（SVD Analysis Transformer SoftMax）算法，该算法实现了对F、G、K型恒星光谱的自动分类。SATS算法，首先以奇异值分解（SVD）的方式，对归一化后的恒星光谱做去噪处理；然后对恒星光谱进行特征提取，特征提取层共有六个模块：增量主成分分析（Incremental PCA）、核主成分分析（Kernel PCA）、稀疏主成分分析（Sparse PCA）、因子分析（Factor Analysis）、独立成分分析（Fast ICA）和Transformer（前五个模块统称为Analysis模块），为保证方差贡献率在0.95以上，IPCA、KPCA、Sparse PCA、Factor Analysis和Fast ICA将恒星光谱特征提取为300维；最后，将恒星光谱输入到SoftMax层进行自动分类。SATS算法将多个Analysis模块结合使用，进一步提高了使用单一Analysis模块的分类正确率。Transformer模块和多个Analysis模块的结合使用，再一次提高了分类正确率。SATS算法最大的优势在于对恒星光谱进行了多重特征提取，以不同的特征提取方式，最大程度地保留了原始恒星光谱信息，将信息损失做到最低。SATS算法的最终分类正确率为0.93，这一分类正确率也高于混合深度学习算法CNN(convolutional neural network)+Bayes、CNN+KNN、CNN+SVM、CNN+Adaboost和CNN+Adaboost0.86、0.88、0.89、0.87、0.89的分类正确率。

关键词：Transformer；LAMOST；恒星光谱；SVD；CNN

Abstract：An in-depth study of a stellar's spectrum provides insight into its chemical composition and physical properties. Stellar spectrum classification is an important direction in stellar spectrum research. With the emergence of massive stellar spectrum data, artificial classification cannot meet scientific research needs. Based on this, this paper constructs the SATS algorithm, which realizes the automatic classification of F, G, and K-type stellar spectra. Firstly, the SATS algorithm uses singular value decomposition (SVD) to denoise the normalized stellar spectra. Then, the SATS algorithm performs feature extraction on the stellar spectrum. The feature extraction layer consists of six modules:Incremental principal component analysis (IPCA), nuclear principal component analysis (KPCA), sparse principal component analysis (SparsePCA), FactorAnalysis, independent component analysis (FastICA) and Transformer(the six modules are collectively referred to as Analysis module), to ensure that the variance contribution rate is above 0.95, IPCA, KPCA, SparsePCA, FactorAnalysis and FastICA extract the stellar spectral features into 300 dimensions. Finally, the stellar spectra are fed into the SoftMax layer for automatic classification. SATS algorithm combines multiple analysis modules to improve the accuracy of classification further using a single analysis module. Once again, the combination of Transformer modules and multiple Analysis modules improves classification accuracy. The most significant advantage of the SATS algorithm is that it performs multiple feature extraction on the stellar spectrum, which retains the stellar spectral information to the maximum extent and minimizes the information loss by different feature extraction methods. The final classification accuracy of the SATS algorithm is 0.93, which classification accuracy is also higher than that of the hybrid deep learning algorithms CNN+Bayes, CNN+Knn, CNN+SVM, CNN+Adaboost and CNN+Adaboost 0.86, 0.88, 0.89, 0.87, 0.89.

Key words：Transformer; LAMOST; Stellar spectra; SVD; CNN

收稿日期: 2022-11-22 修订日期: 2023-10-02

中图分类号:

P157.2

基金资助: 国家自然科学基金天文联合基金项目(U1731128)资助

通讯作者: 李建喜 E-mail: ptjxli@hotmail.com

作者简介: 屠良平，1977年生，辽宁科技大学理学院教授 e-mail: tlp_kd@163.com

引用本文:

屠良平，李双川，涂东鑫，李建喜，丁治超. SATS：一种基于多重特征提取的恒星光谱分类算法[J]. 光谱学与光谱分析, 2024, 44(07): 2029-2036.
TU Liang-ping, LI Shuang-chuan, TU Dong-xin, LI Jian-xi, DING Zhi-chao. SATS: A Stellar Spectral Classification Algorithm Based on Multiple Feature Extraction. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2024, 44(07): 2029-2036.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2024)07-2029-08 或 https://www.gpxygpfx.com/CN/Y2024/V44/I07/2029