针对LAMOST DR5 pipeline分类为Unknown的光谱数据对其进行了特征提取和聚类分析。 主要工作如下: (1) 基于影响空间及数据场的特征提取。 首先基于影响空间从低信噪比光谱中提取出大量小集团; 然后计算各小集团内部的场并根据场对光谱排序, 依次访问光谱序列及其小集团内的成员来获得特征谱; (2) 对上述特征谱进行K-means聚类, 并统计了每一类目标所在天区、 观测视宁度、 各波段信噪比、 亮度、 光谱仪/光纤的分布情况。 (3) 低质量光谱聚类结果的理论分析。 通过聚类所有低质量光谱被分为了5大簇: A 光谱信噪比较低或与传统分类模板差异较大, 但通过特征分析可确定其类别(占比2.7%); B 光谱蓝端或红端出现疑似特征线或分子带, 但与线表无法匹配(占比23.6%); C 光谱蓝端信噪比极低, 且该波长区域噪声值较强, 其他波长区域的连续谱和线的特征较弱(占比48.0%); D 红蓝两端拼接问题导致5 700~5 900 Å局部光谱突起明显, 其他波长区域的连续谱和线的特征较弱(占比24.2%); E 存在大量缺省值导致无法确定其类别(占比1.5%)。 实验结果表明, 该方法不仅能够有效提取低信噪比光谱的特征谱, 同时能够通过特征谱的聚类分析揭示低质量光谱的成因, 从而为制定光谱观测计划提供参考, 为低信噪比光谱分析及处理提供方法借鉴。

Based on the spectral data classified as Unknown by LAMOST DR5 Pipeline, the characteristics of low-quality spectra are extracted, and clustering analysis is conducted in this paper. The main work includes: (1) Feature extraction based on influence space and the data field. Firstly, a large number of small clusters are extracted from the low SNR spectrum based on influence space; secondly, each small cluster’s data field is calculated, and the spectrum is sorted using the above field; and then, access the sorted spectrum and the members in its small cluster to obtain the characteristic spectrum. (2) Carry out K-means clustering with the above characteristic spectrum and statistics on the sky area, observed visual ninety, the signal-to-noise ratio in each band, brightness, and spectrometer/fiber distribution for each class of targets. (3) Analysis of clustering results of the low SNR spectra. All low-quality spectra are divided into five categories through cluster analysis: A. The spectral SNR is low, or the spectrum is different from the traditional classification template, but its category can be determined by feature analysis (accounting for 2.7%); B. Suspected characteristic lines or molecular bands that do not match the line table appear at the blue or red end of the spectrum (accounting for 23.6%); C. The SNR at the spectrum’s blue end is very low, and the noise value in this wavelength region is strong. While in other wavelength regions, the features of continuous spectrum and line are weak (accounting for 48%); D. Due to the splicing problem, a protrusion can be seen in the local spectrum between 5 700 and 5 900 Å, and the continuum and line characteristics are poor at other wavelengths (accounting for 24.2%); E. Many default values make it impossible to determine the category of the spectrum (accounting for 1.5%). The experimental results show that this method can not only effectively extract the characteristic spectrum of low SNR spectrum, but also effectively carry out clustering analysis on the characteristic spectrum to reveal their causes, to provide a reference for the formulation of spectrum observation plan and the analysis and processing of low SNR spectrum.