中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
恒星光谱大气物理参量自动估计研究

文献类型:学位论文

作者张健楠
学位类别工学博士
答辩日期2005-12-18
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
关键词恒星光谱 表面有效温度 表面重力加速度 化学丰度 PCA 非参数回归 偏最小二乘回归(PLSR) 核主成分回归(KPCR) 最小二乘回归(LSR) Stellar Spectra Effective Temperature Gravity Metallic principal component analysis (PCA) Non-parameter regression (NPR) Partial least squared regression (PLSR) kernel PCA regression (KPCA) Least squares regression (LSR)
其他题名Studies on Automated Measurement of Stellar Spectral Fundamental Parameters
学位专业模式识别与智能系统
中文摘要本文主要讨论视向速度已作校正的低分辨率恒星光谱大气基本物理参量(表面有效温度(Teff),表面重力(log g),化学丰度([M/H]))的自动估计问题。将物理参量的自动估计视为从光谱到其物理参量的回归问题。提出了自动估计物理参量的四种回归模型,并进行了大量的实验比较研究。 a) 恒星表面有效温度的曲面拟合方法 针对连续温度分布的不同恒星光谱数据在前两个主成分空间的分布规律,提出了一种估计有效温度的曲面拟合方法。曲面模型是以10为底的指函数与多项式函数的复合模型。实验结果表明:使用3次多项式的复合模型所得到的拟合曲面,不仅有较好的拟合精度而且有很好的鲁棒性。 b) 恒星物理参量的非参数回归方法 提出两种确定恒星物理参量的非参数回归模型,一种是基于固定核宽的非参数回归,另一种是基于变核宽的非参数回归。在固定核宽方法中,选择合适的核宽是较为困难的。但是,在变核宽方法中,核宽随样本点的分布密度而发生变化,由模型本身确定。实验结果表明可变核宽方法与固定核宽方法相比,物理参量估计的精度和鲁棒性更高。 c) 恒星物理参量的偏最小线性二乘回归技术(PLSR) 研究了线性回归模型的PLSR算法在低分辨率恒星光谱基本物理参量估计中的应用。提出了分段PLSR方法,提高了PLSR方法关于物理参量的估计效果。实验结果表明分段PLSR方法,对表面重力和化学丰度的估计都优于其它方法,如人工神经网络方法、K-近邻方法和非参数回归方法等。 d) 恒星基本物理参量的非线性核回归技术 研究了非线性核回归模型在低分辨率恒星光谱基本物理参量估计中的应用。采用核方法将线性回归模型扩展为非线性回归模型,包括核最小二乘回归(KLSR),核PCA回归(KPCR),核偏最小二乘回归(KPLSR)。实验结果表明:对于温度参数估计,三种算法具有相近的估计效果;对于表面重力和化学丰度估计,KPCR和KPLSR具有相近的估计效果,但它们都优于KLSR和非参数回归方法;当数据噪声较大时,KPCR,KPLSR和非参数回归三者有相近的估计效果。
英文摘要This work is focused on automated measuring stellar fundamental parameters, effective temperature (Teff), gravity (log g), and metallic ([M/H]), from low-resolution spectra which radial velocities have been calibrated. The procedure of measuring the three parameters from spectral data is regarded as a regression problem. Four regression methods for the stellar parameter estimation are proposed, and those methods are tested in extensive experiments. a) Surface fitting method for the effective temperature estimation. Based on the variation of the two-dimensional PCA projections of stellar spectral data with effective temperature, a surface fitting method to estimate the effective temperature is proposed. The surface model is a composition of the exponential function to base 10 and a polynomial. The experiments show that using a cubic polynomial can achieve better performance for estimating the effective temperature. b) Non-parameter regression (NPR) for the stellar parameter estimation. Two NPR models for the stellar parameter estimation are proposed. One is NPR with a fixed kernel width; the other is with variable kernel width. Selecting an appropriate kernel width is difficult in the fixed kernel width method. In the variable kernel width method, the kernel widths are self-adaptive, and are varied with the densities of samples. Experiments show that the variable kernel width method is more efficient and robust than the fixed one. c) Partial linear squares regression (PLSR) for the stellar parameter estimation. The PLSR is applied in the stellar parameter estimation from low-resolution spectra. A piecewise PLSR method is presented. It effectively boosts up the performance of the PLSR for the stellar parameter estimation. The experiments for gravity and metallic estimation show that the piecewise PLSR method is superior to some other methods such as the ANN, the k-near neighbours and the NPR. d) Non-linear regression based on kernel for the stellar parameter estimation. The linear regression model is extended to non-linear regression by using a kernel function for the stellar parameter estimation from low-resolution spectra. Three non-linear regression algorithms, the kernel least squared regression (KLSR), the kernel PCA regression (KPCR) and the kernel PLSR (KPLSR), are investigated. Extensive experiments show: 1). For the effective temperature estimation, the three algorithms perform similarly; 2). For the gravity and metallic estimation, the KPCR and the KPLSR perform similarly, and the both of them are superior to the KLSR and the NPR;3). With lower signal-to-noise ratio, the KPCR,the KPLSR and the NPR perform similarly for the stellar parameter estimation.
语种中文
其他标识符200218014603233
源URL[http://ir.ia.ac.cn/handle/173211/5883]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
张健楠. 恒星光谱大气物理参量自动估计研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2005.

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。