星系光谱的自动识别与分类技术研究
文献类型：学位论文
| 作者 | 赵梅芳 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2006-05-27 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 吴福朝 ; 赵永恒 |
| 关键词 | 天体光谱 自动分类 径向基神经网络 动态衰减调节 Adaboost 特征融合 Celestial object spectra automated classification radial basis function neural network dynamic decay adjustment Adaboost feature fusion |
| 其他题名 | Automated spectral recognition and classification of galaxies |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 我国目前正在建造世界上最大规模、总投资达2.35亿人民币的大型天文观测仪器“大天区面积多目标光纤光谱望远镜” (LAMOST)。建成后,将成为世界上最大视场(5度)兼最大口径(4米)，光谱观测效率最高的光学天文望远镜，每个观测夜将获得1～2万条光谱数据，预计所获得的光谱数据总量达107<上标!>。以往天文学家进行人工识别光谱的方式已不能适用，本文以LAMOST项目为背景，主要探讨星系的的自动识别与分类算法，以满足LAMOST项目的需要。 本文主要工作包括以下四个部分： （1）基于自适应径向基神经网络（Adaptive-RBFN）的类星体自动识别方法 本文提出了自适应径向基神经网络的类星体识别算法，将类星体从所有星系光谱中首先区分出来。Adaptive-RBFN是对神经元进行监督的自适应增加神经元的RBFN算法，实验表明它不但克服了经典RBFN算法选择隐层神经元数目的困难，而且还提高了对类星体识别的稳定性和正确率。实验表明，该方法具有训练步骤简洁，识别结果稳定等优点。 （2）对类星体和星系光谱自动分类的快速神经网络方法 针对已有的神经网络技术中的速度问题，提出了动态衰减调节算法与径向基相结合(RBFN-DDA)的神经网络方法。实验对比了最近邻、分类判别的覆盖算法、RBF方法和自适应增加节点的RBF方法对于星系和类星体的光谱分类性能。此方法具有拒绝决策，实验结果表明，RBFN-DDA方法的错误识别率最低，拒识的样本数目相对比较少，可以通过其他方法再进一步识别和分类，所以此方法的实验结果的置信度比其它几种方法要高。此方法运算速度很快，适合大规模的光谱数据的快速处理。 （3）活动星系核（AGNs）与星暴星系（Starburst Galaxies）的自动识别研究 结合以星暴星系为主的恒星形成星系和活动星系核在光谱中的发射线的不同表征，在恢复到零红移状态后的光谱上截取有效发射线波段范围，采用经典的K近邻方法，对它们进行自动分类研究。星暴星系和活动星系核光谱中的主要区别在于Hβ、[OIII]、Hα和[NII]等发射线的幅度和半高全宽（FWHM）的大小，所以截取这些发射线所在的波段进行单独或组合的分类实验，实验证明，从整条连续光谱中抽取以Hβ和[OIII]发射线为主的波段和以Hα和[NII]发射线为主的波段直接进行特征合并后，分类的效果最好。经过实验分析，在充分利用光谱的典型特征的情况下，自动分类方法也可有效地应用于活动星系的分类，与传统的通过计算发射线的FWHM值或发射线强比对光谱进行分类的方法相比，此方法不需测量谱线参数，为大型光谱巡天所产生的庞大数据库提供了一种快速直接的自动分类方法。 （4）宽线AGNs与窄线AGNs的自动识别研究 首先，在恢复到零红移状态后的AGNs光谱上截取具有有效特征的波段范围，再采用自适应增强（Adaboost）的方法，对宽线和窄线AGNs进行特征融合的分类实验，确定了以Hα和[NII]发射线为主的波段为宽线和窄线AGNs光谱的主要区别特征。然后，单独对Hα和[NII]发射线为主的波段，用自适应增强的方法进行光谱自动分类。自适应增强方法方法不需要事先调节参数，且“弱分类器”的分类结果只需好于随机猜测，算法简单。实验证明，对于单独采用以Hα和[NII]发射线为主的波段，自适应增强方法能达到较好的分类效果，从而可有效地应用于大型光谱巡天所产生的活动星系核光谱的自动分类中。 |
| 英文摘要 | The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) project is one of the National Major Scientific Projects undertaken by the Chinese Academy of Science. The total expected number of spectra to be acquired is about 107<上标!>. This work is particularly focused on finding and designing suitable techniques for automatic spectral classification and recognition. The main contributions are following: (1) QSOs Spectral recognition technique based on Adaptive Radial Basis Function Neural Network This paper presents a novel adaptive method based on the radial basis function neural networks (RBFN) to recognize Quasars out of spectra of galaxies. The Adaptive-RBFN adaptively increases the number of neurons in the hidden layer. Experiments show that our method can not only overcome the difficulty of selecting the number of neurons in hidden layer of the traditional RBFN algorithm, but also increase the stability and accuracy of quasars recognition. Besides its simple training steps and can acquire stable recognition results. (2) Spectral classification of QSOs and galaxies based on fast neural network A fast neural network method based on radial basis function with dynamic decay adjustment (RBF-DDA) is proposed and compared with other methods. Experiments show that the RBFN-DDA has rejection decision-making, and can achieve lower error rates of recognition and get few rejected samples that are susceptible to be later recognized by other methods. So the confidence level of this method is higher. The RBFN-DDA runs fast and fit for the processing of large-scale spectra data. (3) A study on the automated classification of AGNs and Starburst Galaxies Based on the truncated effective emission-line wave bands from the spectra in rest frame, the K-nearest-neighbour method is applied for the automated recognition. It is shown by experiments that when the wave band of Hβ and [OIII] and that of Hα and [NII] are concatenated directly, the best results are achieved. The analysis of the experiments shows by fully exploiting typical spectral features, active galaxies can be classified automatically and effectively. (4) A study on the automated classification of broad-line AGNs and narrow-line AGNs An adaptive boosting (Adaboost) method is developed for the classification by feature fusion, and the wave band of Hα and [NII] is shown to be the main discriminative feature between broad-line AGNs and narrow-line AGNs. The Adaboost method needs not to adjust parameters in advance and the results of “weak classifiers” are only required to be better than random guessing, so its algorithm is very simple. The experiments show that the Adaboost method achieves good classification performance only based on the wave band of Hα and [NII], it could be applied to the automatic classification of large amount of AGN spectra from the large-scale spectral surveys. |
| 语种 | 中文 |
| 其他标识符 | 200318014603044 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/5904]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 赵梅芳. 星系光谱的自动识别与分类技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2006. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。