基于CNN与LSTM复合深度模型的恒星光谱分类算法  

A Stellar Spectrum Classification Algorithm Based on CNN and LSTM Composite Deep Learning Model

作　　者：李浩 赵青[1] 崔辰州[2] 樊东卫[2] 张成奎 史艳翠 王嫄 LI Hao;ZHAO Qing;CUI Chen-zhou;FAN Dong-wei;ZHANG Cheng-kui;SHI Yan-cui;WANG Yuan(School of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457,China;National Astronomical Observatory,Chinese Academy of Sciences,Beijing 100012,China)

机构地区：[1]天津科技大学人工智能学院,天津300457 [2]中国科学院国家天文台,北京100012

出　　处：《光谱学与光谱分析》2024年第6期1668-1675,共8页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金青年科学基金项目(11803022);国家自然科学基金面上项目(12273077);国家重点研发计划项目(2022YFF0711500)资助。

摘　　要：恒星光谱分类是天文学领域中非常重要的研究方向。随着科技的迅猛发展,大型巡天望远镜采集的恒星光谱数据已经达到了TB或甚至PB级别,传统的分类方法已经无法满足如此庞大数据量的处理需求。正确分类光谱对于研究恒星的物理性质以及演化过程具有重要意义。CNN通过卷积运算学习数据的局部特征,去除冗余信息,并通过最大池化运算对特征进行压缩。然而,由于原始CNN模型的全连接层缺乏长距离依赖挖掘的功能,如果加入LSTM网络,通过其独特的三个“门”的远距离依赖挖掘能力可提取的重要特征,并检测特征中的微小差异,恰好可以解决这个问题。因此,提出了一种基于CNN和LSTM复合的深度模型,用于对LAMOST DR8中的恒星光谱进行分类。这种模型能够更好地学习恒星光谱的特征,为恒星演化研究提供了重要的帮助。为了提高模型的收敛速度,使用了常见的Z-Score标准化方法对数据进行处理。提出的模型在F、G、K三分类实验中取得了94.56%的分类准确率。同时,与前人使用过的RBM、PILDNN、PILDNN、DBN、Inception v3、1D-SSCNN、LSTM方法进行对比,结果表明该方法具有更高的分类准确率。在十分类实验中,该方法取得了97.35%的准确率,并且相比于仅使用LSTM、1D-SSCNN方法的实验结果,该方法的结果更好,且训练时间减少了近十倍。使用F1-score对每类恒星光谱分类准确度进行说明,在三分类和十分类实验中,每类的F1值都在0.9以上。与前人在文献中的实验结果进行对比,该模型的结果更好。通过混淆矩阵的结果,可以得出该模型在光谱种类越多的实验中准确率越高,甚至可以达到100%。综上所述,所提出的基于CNN和LSTM相结合的模型可以有效地对大规模恒星光谱数据进行分类,并取得了优异的分类效果。Stellar spectral classification is a significant research direction in astronomy.With the rapid development of technology,the stellar spectral data collected by large survey telescopes have reached terabytes or even petabytes,and the traditional classification methods can no longer meet the processing needs of such a vast amount of data.CNNs learn the local features of the data by convolution operations,remove redundant information,and compress the features by maximum pooling operations.However,since the fully-connected layer of the original CNN model lacks the function of long-range dependency mining,this problem can be solved by adding LSTM networks,which can extract important features and detect small differences in features through their unique three“gates”of long-range dependency mining capability.Therefore,this paper proposes a deep model based on the composite of CNN and LSTM for classifying stellar spectra in LAMOST DR8.This model can better learn the features of stellar spectra,which provides an important help for stellar evolution studies.To improve the convergence speed of the model,the common Z-Score normalization method is used to process the data.The model proposed in this paper achieved a classification accuracy of 94.56%in the F,G,and K classification experiments.Meanwhile,compared with the previously used RBM,PILDNN,PILDNN,DBN,Inception v3,1D-SSCNN,and LSTM methods,the results show that the method in this paper has a higher classification accuracy.In the ten-class experiments,the method in this paper achieves 97.35%accuracy.The results are better than the experimental results using only LSTM and 1D-SSCNN methods,and the training time is reduced by nearly ten times.The F1 score is used to illustrate the classification accuracy of each class of stellar spectra,and the F1 value of each type is above 0.9 in both the three-classification and ten-class experiments.Compared with the results of previous experiments in the literature,the results of this paper's model are better.With the confusion matr

关 键 词：LAMOST 恒星光谱分类 卷积神经网络 长短期记忆网络 Z-Score标准化 

分 类 号：P157.2[天文地球—天文学]