基于Stacking集成学习的恒星/星系分类研究

机器学习在当今诸多领域已经取得了巨大的成功,但是机器学习的预测效果往往依赖于具体问题.集成学习通过综合多个基分类器来预测结果,因此,其适应各种场景的能力较强,分类准确率较高.基于斯隆数字巡天(Sloan Digital Sky Survey,SDSS)计划恒星/星系中最暗源星等集分类正确率低的问题,提出一种基于Stacking集成学习的恒星/星系分类算法.从SDSS-DR7(SDSS Data Release 7)中获取完整的测光数据集,并根据星等值划分为亮源星等集、暗源星等集和最暗源星等集.仅针对分类较为复杂且困难的最暗源星等集展开分类研究.首先,对最暗源星等集使用10折嵌套交叉验证,然后使用支持向量机(Support Vector Machine,SVM)、随机森林(Random Forest,RF)、XGBoost(eXtreme Gradient Boosting)等算法建立基分类器模型;使用梯度提升树(Gradient Boosting Decision Tree,GBDT)作为元分类器模型.最后,使用基于星系的分类正确率等指标,与功能树(Function Tree,FT)、SVM、RF、GBDT、XGBoost、堆叠降噪自编码(Stacked Denoising AutoEncoders,SDAE)、深度置信网络(Deep Belief Network,DBN)、深度感知决策树(Deep Perception Decision Tree,DPDT)等模型进行分类结果对比分析.实验结果表明,Stacking集成学习模型在最暗源星等集分类中要比FT算法的星系分类正确率提高了将近10%.同其他传统的机器学习算法、较强的提升算法、深度学习算法相比,Stacking集成学习模型也有较大的提升.     

TESS空间卫星单扇区变星的周期性研究

TESS (Transiting Exoplanet Survey Satellite)空间卫星提供的短曝光、高精度光度测量为寻找并区分变星与搜寻行星提供了良好的数据.利用变星源的光变曲线,使用周期频谱分析与光变折叠等一系列方法分析了TESS空间卫星21扇区19995颗拥有高质量光变数据的目标源,并对这些源进行了分类,共获得4624颗变星,其中食双星322颗、脉动变星470颗、行星凌星37颗.所得变星结果与VSX (The International Variable Star Index)变星表进行了交叉比较,共交叉匹配了625颗变星源,这些交叉源中共有131颗为食双星系统、31颗为脉动变星,并通过周期频谱分析获取了双星绕转以及脉动周期.另外在59颗变星中发现明显耀发现象,交叉源中有8颗变星为行星凌星并同样通过周期频谱分析获取了行星绕转周期,从而验证了TESS空间卫星数据对变星分析的可行性.通过利用TESS空间卫星21扇区获得的变星周期结果与VSX变星表中提供的变星周期对比,发现与VSX变星表中绝大部分变星的周期一致,有一部分结果与VSX变星表中的结果差别较大,对这些变星周期结果做了进一步修正,并给出了变星表未列出的变星周期结果.     

EfficientNetV2-S-Triplet7: 一种改进的星系newline 形态学分类算法

星系的形态与星系的形成和演化息息相关, 其形态学分类是星系天文学后续研究的重要一环. 当前海量天文观测数据的出现使得天文数据自动分析方法越来越得到重视, 针对此问题, 利用先进的深度学习骨干网络EfficientNetV2, 分析不同的注意力机制类型和使用节点对网络性能的影响, 构建了一种命名为EfficientNetV2-S-Triplet7 (即在EfficientNetV2-S stage7的$1\times1$卷积层后加入Triplet模块)的改进算法模型来实现星系形态学的自动分类. 使用第二期星系动物园(Galaxy Zoo 2, GZ2)中超过24万张的测光图像作为初始数据进行实验测试. 在对数据进行预处理时采取了尺寸抖动、翻转、色彩畸变等图像增强手段来解决图像数量的不平衡问题. 在同一系列经典和前沿的深度学习算法模型AlexNet、ResNet-34、MobileNetV2、RegNet进行对比实验后, 得出EfficientNetV2-S-Triplet7算法在分类准确率、查全率和F1分数等指标上具有最好的测试结果. 在9375张测试图像中的3项指标值分别可达到89.03%、90.21%、89.93%, 查准率达到89.69%, 在其他模型中排在第3位. 该结果表明将EfficientNetV2-S-Triplet7算法应用于大规模星系数据的形态学分类任务中有很好的效果.     

基于EfficientNet的星系形态分类研究

星系的结构和形态能够反映星系自身的物理性质,其形态的分类是后续分析研究的一个重要环节.EfficientNet模型使用复合系数对深度网络模型的深度、宽度、输入图像分辨率进行更加结构化的统一缩放,是一种新的深度网络优化扩展方法.将该模型应用于星系数据形态的分类研究中,结果表明基于EfficientNetB5模型的平均准确率、精确率、召回率以及F1分数(精确率与召回率的调和平均数)都在96.6%以上,与残差网络(Residual network, ResNet)中ResNet-26模型的分类结果相比有较大的提升.实验结果证明EfficientNet的深度网络优化扩展方法可行且有效,可应用于星系的形态分类.     

一种新的基于2维傅里叶谱图像的恒星光谱特征提取方法和深度网络分类应用

天体光谱分类是天文学研究的重要内容之一,其关键是从光谱数据中选择和提取对分类识别最有效的特征构建特征空间.提出一种新的基于2维傅里叶谱图像的特征提取方法,并应用于LAMOST (the Large Sky Area Multi-Object Fiber Spectroscopic Telescope)恒星光谱数据的分类研究中.光谱数据来源于LAMOST Data Release 5(DR5),选取30000条F、 G和K型星光谱数据,利用短时傅里叶变换(Short-Time Fourier Transform, STFT)将1维光谱数据变换成2维傅里叶谱图像,对得到的2维傅里叶谱图像采用深度卷积网络模型进行分类,得到的分类准确率是92.90%.实验结果表明通过对LAMOST恒星光谱数据进行STFT可得到光谱的2维傅里叶谱图像,谱图像构成了新的光谱数据特征和特征空间,新的特征对于光谱数据分类是有效的.此方法是对光谱分类的一种全新尝试,对海量天体光谱的分类和挖掘处理有一定的开创意义.     

基于卷积神经网络的恒星光谱型和光度型的分类模型

恒星光谱分类是天文学中一个重要的研究问题.对于已经采集到的海量高维恒星光谱数据的分类,采用模式匹配方法对光谱型分类较为成功,但其缺点在于标准恒星模版之间的差异性在匹配实际观测数据中不能体现出来,尤其是当需要进行光谱型和光度型的二元分类时模版匹配法往往会失败.而采用谱线特征测量的光度型分类强烈地依赖谱线拟合的准确性.为了解决二元分类的问题,介绍了一种基于卷积神经网络的恒星光谱型和光度型分类模型(Classification model of Stellar Spectral type and Luminosity type based on Convolution Neural Network, CSSL CNN).这一模型使用卷积神经网络来提取光谱的特征,通过注意力模块学习到了重要的光谱特征,借助池化操作降低了光谱的维度并压缩了模型参数的数量,使用全连接层来学习特征并对恒星光谱进行分类.实验中使用了大天区面积多目标光纤光谱天文望远镜(Large Sky Area Multi-Object Fiber Spectroscopy Telescope, LAMOST)公开数据集Data Release 5 (DR5,用了其中71282条恒星光谱数据,每条光谱包含了3000多维的特征)对该模型的性能进行验证与评估.实验结果表明,基于卷积神经网络的模型在恒星的光谱型分类上准确率达到92.04%,而基于深度神经网络的模型(Celestial bodies Spectral Classification Model, CSC Model)只有87.54%的准确率; CSSL CNN在恒星的光谱型和光度型二元分类上准确率达到83.91%,而模式匹配方法MKCLASS仅有38.38%的准确率且效率较低.     

基于SE-Inception-v3的星系形态分类模型

随着天文探测技术的快速发展, 海量的星系图像数据不断产生, 能够及时高效地对星系图像进行形态分类对研究星系的形成与演化至关重要. 针对传统的星系形态分类模型特征选择困难、分类速度慢、准确率受限等难题, 提出一种以Inception-v3神经网络为主干结构, 融合压缩激励(Squeeze and Excitation Network, SE)通道注意力机制的星系形态分类模型. 该模型在斯隆数字巡天(Sloan Digital Sky Survey, SDSS)样本的测试集准确率高达99.37%. 旋涡星系、圆形星系、中间星系、雪茄状星系与侧向星系的F1值分别为99.33%、99.58%、99.33%、99.41%与99.16%. 该模型与Inception-v3、MobileNet (Mobile Neural Network)和ResNet (Residual Neural Network)网络模型相比, SE-Inception-v3宽度和深度优势表现出更强的特征提取能力, 可以高效识别不同形态的星系, 为未来大型巡天计划的大规模星系形态分类问题提供了一种新方法.     

基于先验信息的空间碎片探测方法

提出了一种基于先验信息的空间碎片图像探测方法.该方法通过先验信息,在图像中碎片星像的邻域设置波门,计算波门内的局部背景阈值,辅以相关的判据识别目标.随后使用矩方法计算碎片质心相对波门中心的偏离值,通过线性平移计算碎片质心在整幅图像中的位置.实验表明:该方法复杂度低、便于实现、实时性好,可以高效、准确地探测空间碎片,比较精确地确定碎片的质心位置.     

Blazar天体3C 66A光学波段准周期光变分析

研究了Blazar天体3C 66A光学波段的准周期光变行为.收集了3C 66A光学V波段将近18 yr (2003—2021年)的测光数据,观测数据主要来源是:上海天文台(ShAO)、 AAVSO (The American Association of Variable Star Observers)数据库、Steward天文台.使用了Jurkevich和Lomb-Scargle两种方法分析了光变数据.Jurkevich方法得到了(850±90) d (～2.3 yr)和(1150±140) d (～3.2 yr)的光变周期,而Lomb-Scargle方法在充分考虑了“红噪声”效应之后同样得到了(869±70) d和(1111±90) d的光变周期,它们的置信水平分别为>99%和> 95%.通过与之前的研究结果比较,发现～2.3 yr的光变周期在3C 66A的历史光变数据中是一个稳定的周期,而～3.2 yr的周期则是不稳定的.     

卫星轨道确定模型误差的深度加权核估计方法

卫星动力学模型误差是客观存在的事实,动力学模型误差传递到轨道确定算法中构成部分形式未知的模型误差,并且与测量系统自身的系统误差和随机误差耦合在一起形成定轨模型误差,严重影响轨道确定精度.详细推导了存在动力学模型误差的轨道改进方程,对模型中能准确描述的部分建立了参数化模型,对不能准确描述的误差部分,建立了非参数模型.构建了部分线性轨道改进模型,利用二阶段估计法和核函数估计法对模型误差进行拟合估计,并在轨道改进中予以补偿.根据数据深度理论,建立了非参数模型误差的深度加权核估计方法,提高了模型误差估计的抗差性.最后结合天基空间目标监视系统进行了轨道确定仿真实验.实验结果表明,模型误差是影响轨道确定精度的重要因素,核函数估计法可以有效估计定轨中的模型误差,窗宽是提高模型估计精度的重要变量,通过深度加权处理可以明显提高核函数估计的抗差性,提高轨道确定精度.     

The Basic Shape Classification of Space Debris with Light Curves

We study the machine learning method for classifying the basic shape of space debris in both simulated and observed data experiments, where light curves are used as the input features. In the dataset for training and testing, simulated light curves are derived from four types of debris within different shapes and materials. Observed light curves are extracted from Mini-Mega TORTORA (MMT) database which is a publicly accessible source of space object photometric records. The experiments employ the deep convolutional neural network, make comparisons with other machine learning algorithms, and the results show CNN (Convolutional Neural Network) is better. In simulational experiments, both types of cylinder can be distinguished perfectly, and two other types of satellite have around 90% probability to be classified. Rockets and defunct satellites can achieve 99% success rate in binary classification, but in further sub-classes classifications, the rate becomes relatively lower.     

Detrending time series for astronomical variability surveys

We present a detrending algorithm for the removal of trends in time series. Trends in time series could be caused by various systematic and random noise sources such as cloud passages, changes of airmass, telescope vibration, CCD noise or defects of photometry. Those trends undermine the intrinsic signals of stars and should be removed. We determine the trends from subsets of stars that are highly correlated among themselves. These subsets are selected based on a hierarchical tree clustering algorithm. A bottom-up merging algorithm based on the departure from normal distribution in the correlation is developed to identify subsets, which we call clusters. After identification of clusters, we determine a trend per cluster by weighted sum of normalized light curves. We then use quadratic programming to detrend all individual light curves based on these determined trends. Experimental results with synthetic light curves containing artificial trends and events are presented. Results from other detrending methods are also compared. The developed algorithm can be applied to time series for trend removal in both narrow and wide field astronomy.     

空间碎片轨道预报中动力学模型基准的选取

在不同的轨道预报场景中, 使用的动力学模型也不同. 例如, 在低轨空间碎片的预报中大气阻力是十分重要的摄动力, 而到了中高轨, 大气阻力就可以忽略不计. 如何为不同轨道类型的空间碎片选择最优(满足精度要求下的最简)动力学模型还没有系统、详尽的研究. 将对不同精度需求、不同轨道类型下的大批量轨道进行预报, 通过比较不同动力学模型下的预报结果, 给出各种预报场景的最优动力学模型建议. 可以为不同轨道类型的空间碎片在轨道预报时选择基准动力学模型提供参考或标准.     

Comparison between Windowed FFT and Hilbert-Huang Transform for Analyzing Time Series with Poissonian Fluctuations: A Case Study

Hilbert-Huang Transform (HHT) is a novel data analysis technique for nonlinear and non-stationary data. We present a time-frequency analysis of both simulated light curves and an X-ray burst from the X-ray burster 4U 1702-429 with both the HHT and the Windowed Fast Fourier Transform (WFFT) methods. Our results show that the HHT method has failed in all cases for light curves with Poissonian fluctuations which are typical for all photon counting instruments used in astronomy, whereas the WFFT method can sensitively detect the periodic signals in the presence of Poissonian fluctuations; the only drawback of the WFFT method is that it cannot detect sharp frequency variations accurately.     

Finding outlier light curves in catalogues of periodic variable stars

We present a methodology to discover outliers in catalogues of periodic light curves. We use a cross-correlation as the measure of 'similarity' between two individual light curves, and then classify light curves with lowest average 'similarity' as outliers. We performed the analysis on catalogues of periodic variable stars of known type from the MACHO and OGLE projects. This analysis was carried out in Fourier space and we established that our method correctly identifies light curves that do not belong to those catalogues as outliers. We show how an approximation to this method, carried out in real space, can scale to large data sets that will be available in the near future such as those anticipated from the Panoramic Survey Telescope & Rapid Response System (Pan-STARRS) and Large Synoptic Survey Telescope (LSST).     

Are the missing X-ray breaks in gamma-ray burst afterglow light curves merely hidden?

Gamma-ray burst (GRB) afterglow observations in the Swift era have a perceived lack of achromatic jet breaks compared to the BeppoSAX or pre- Swift era. Specifically, relatively few breaks, consistent with jet breaks, are observed in the X-ray light curves of these bursts. If these breaks are truly missing, it has serious consequences on the interpretation of GRB jet collimation and energy requirements, and the use of GRBs as cosmological tools. Here, we address the issue of X-ray breaks that are possibly 'hidden' and hence the light curves are misinterpreted as being single power laws. We do so by synthesizing X-ray telescope (XRT) light curves and fitting both single and broken power laws, and comparing the relative goodness of each fit via Monte Carlo analysis. Even with the well-sampled light curves of the Swift era, these breaks may be left misidentified, hence caution is required when making definite statements on the absence of achromatic breaks.     

Planetesimals to protoplanets – II. Effect of debris on terrestrial planet formation

In this paper, we extend our numerical method for simulating terrestrial planet formation to include dynamical friction from the unresolved debris component. In the previous work, we implemented a rubble pile planetesimal collision model into direct N -body simulations of terrestrial planet formation. The new collision model treated both accretion and erosion of planetesimals but did not include dynamical friction from debris particles smaller than the resolution limit for the simulation. By extending our numerical model to include dynamical friction from the unresolved debris, we can simulate the dynamical effect of debris produced during collisions and can also investigate the effect of initial debris mass on terrestrial planet formation. We find that significant initial debris mass, 10 per cent or more of the total disc mass, changes the mode of planetesimal growth. Specifically, planetesimals in this situation do not go through a runaway growth phase. Instead, they grow concurrently, similar to oligarchic growth. The dynamical friction from the unresolved debris damps the eccentricities of the planetesimals, reducing the mean impact speeds and causing all collisions to result in merging with no mass loss. As a result, there is no debris production. The mass in debris slowly decreases with time. In addition to including the dynamical friction from the unresolved debris, we have implemented particle tracking as a proxy for monitoring compositional mixing. Although there is much less mixing due to collisions and gravitational scattering when dynamical friction of the background debris is included, there is significant inward migration of the largest protoplanets in the most extreme initial conditions (for which the initial mass in unresolved debris is at least equal to the mass in resolved planetesimals).     

A search for planets transiting the M-dwarf debris disc host, AU Microscopii

We present high-cadence, high-precision multiband photometry of the young, M1Ve, debris disc star, AU Microscopii. The data were obtained in three continuum filters spanning a wavelength range from 4500 to 6600 Å, plus Hα, over 28 nights in 2005. The light curves show intrinsic stellar variability due to star-spots with an amplitude in the blue band of 0.051 mag and a period of 4.847 d. In addition, three large flares were detected in the data which all occur near the minimum brightness of the star. We remove the intrinsic stellar variability and combine the light curves of all the filters in order to search for transits by possible planetary companions orbiting in the plane of the nearly edge-on debris disc. The combined final light curve has a sampling of 0.35 min and a standard deviation of 6.8 mmag. We performed Monte Carlo simulations by adding fake transits to the observed light curve and find with 95 per cent significance that there are no Jupiter mass planets orbiting in the plane of the debris disc on circular orbits with periods,   P ≤ 5  d. In addition, there are no young Neptune like planets (with radii 2.5 times smaller than the young Jupiter) on circular orbits with periods,   P ≤ 3  d.     

Research on Star/Galaxy Classification Based on Stacking Ensemble Learning

Machine learning has achieved great success in many areas today, but the forecast effect of machine learning often depends on the specific problem. An ensemble learning forecasts results by combining multiple base classifiers. Therefore, its ability to adapt to various scenarios is strong, and the classification accuracy is high. In response to the low classification accuracy of the darkest source magnitude set of stars/galaxies in the Sloan Digital Sky Survey (SDSS), a star/galaxy classification algorithm based on the stacking ensemble learning is proposed in this paper. The complete photometric data set is obtained from the SDSS Data Release (DR) 7, and divided into the bright source magnitude set, dark source magnitude set, and darkest source magnitude set according to the stellar magnitude. Firstly, the 10-fold nested cross-validation method is used for the darkest source magnitude set, then the Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) algorithms are used to establish the base-classifier model; the Gradient Boosting Decision Tree (GBDT) is used as the meta-classifier model. Finally, based on the classification accuracy of galaxies and other indicators, the classification results are analyzed and compared with the results obtained by the Function Tree (FT), SVM, RF, GBDT, Stacked Denoising Autoencoders (SDAE), Deep Belief Nets (DBN), and Deep Perception Decision Tree (DPDT) models. The experimental results show that the stacking ensemble learning model has improved the classification accuracy of galaxies in the darkest source magnitude set by nearly 10% compared to the function tree algorithm. Compared with other traditional machine learning algorithm, stronger lifting algorithm, and deep learning algorithm, the stacking ensemble learning model also has different degrees of improvement.