基于卷积神经网络的快速射电暴候选体分类

针对目前从海量的快速射电暴(Fast Radio Burst, FRB)候选体中人工筛选FRB事件难以为继的现状,提出了一种基于卷积神经网络(Convolutional Neural Networks, CNN)的FRB候选体分类方法.首先,通过真实的观测数据和仿真FRB组成训练和测试样本集.其次,建立了二输入的深度卷积神经网络模型,并对其进行训练、测试和优化,获取FRB候选体分类器.最后,利用来自脉冲星的单脉冲数据对该分类器的有效性和性能进行了验证.实验结果表明,该方法可以快速从大量候选体中准确识别出单脉冲事件,极大地提高了FRB候选体的处理速率和效率.     

脉冲星数据比对分析和可视化系统设计与实现

脉冲星数据比对分析和可视化系统(PSRDB,URL:http://www.psrdb.net/),由FAST(Five-hundred-meter Aperture Spherical Radio Telescope)早期科学数据中心团队为快速开展脉冲星候选体比对分析和数据管理研发.通过前端数据提交页面,接收和维护来自FAST及其他研究机构的候选体数据.目前,PSRDB已收录自1967年人类发现第1颗脉冲星以来所有公开文献发表的2811颗脉冲星样本,并采集了当前主要巡天项目尚未正式发表的源和候选体,如FAST多科学目标同时扫描巡天(CRAFTS)候选体数据.基于入库基础数据,利用位置、周期、色散等参数进行比对分析,辅助科研工作者在线检索匹配已知星表数据,最后将检索匹配、比对分析结果生成图表供进一步分析.目前,PSRDB已被应用于FAST脉冲星搜寻和候选体数据管理.未来,PSRDB可在新源认证、后随观测、观测计划制定和原始数据处理流程设计等方面提供数据和工具支撑.     

大规模脉冲星候选体信号的无监督聚类分析研究

随着500 m口径球面射电望远镜(Five-hundred-meter Aperture Spherical radio Telescope, FAST)等大型射电望远镜的建设和使用,脉冲星巡天数据进入PB时代.为解决如此大量高速采样的标量数据挖掘问题,促进新天文现象的发现,提出一种基于无监督聚类的脉冲星候选体筛选方案.该方案采用基于密度层次、划分方法的混合聚类算法,结合MapReduce/Spark并行计算模型和基于滑动窗口的分组策略,进而提高大量候选体信号筛选的效率.通过在脉冲星数据集HTRU2 (High Time Resolution Universe)上的对比实验,结果表明该算法能取得较高的精确度和召回率,分别是0.946和0.905,并且当并行节点足够时,该算法的时间复杂度相比串行执行明显下降.可见,该方法为脉冲星观测大数据的分析挖掘提供一种可行思路.     

基于机器学习的单脉冲搜索候选体识别对FAST观测CRAFTS数据的应用研究

单脉冲搜索作为脉冲星探测的有力工具,在探测旋转射电暂现源以及快速射电暴中扮演着重要角色。为了从海量的射电巡天数据中快速筛选出最有价值的单脉冲搜索候选体,候选体识别已经从早期启发式阈值判断发展到基于机器学习自动识别。对于FAST观测,研究了基于机器学习的单脉冲搜索候选体识别应用到CRAFTS (the commensal radio astronomy FAST survey)超宽带脉冲星数据的性能表现。在评估过程中,使用单脉冲事件组识别(SPEGID)和单脉冲搜索器(SPS)两类自动识别方法,通过7种不同机器学习分类器对CRAFTS基准数据集产生的单脉冲搜索候选体进行自动识别;作为对比,也使用了启发式阈值判断的方法 (RRATtrap和Clusterrank)。结果表明,SPEGID具有最好的性能表现(最高的F1-score值95.1%、次高的召回率95.4%、最低的假阳性率4.7%),SPS具有最快的筛选速度(平均每小时筛选4 010个候选体)。通过对比分析结果,探讨了如何基于FAST观测数据开展高效的单脉冲搜索候选体识别。     

人工智能在脉冲星候选体筛选中的应用

脉冲星搜寻是对脉冲星、引力波,以及对快速射电暴(Fast Radio Burst,简称FRB)等暂现源进行研究的基础。搜寻不仅可以扩大脉冲星样本,还可以发现极端性质的致密星。这有助于研究致密天体状态方程、星际介质、脉冲星导航、引力波探测等课题。目前,射电望远镜的单次巡天就可以产生百万数量级的脉冲星候选体。面对这些海量数据,仅仅依赖人工识别筛选,已不能满足数据的时效需求,更不能实现数据的实时处理。机器学习、计算机视觉应用等人工智能技术自诞生以来,其理论和技术已日益发展成熟,并已成功运用到脉冲星候选体筛选等射电天文研究领域。首先将介绍现有脉冲星搜寻的人工智能方法,再统计和分析已有脉冲星候选体筛选方法的性能,最后对FAST脉冲星候选体筛选工作进行展望。     

基于NPSVM模型的超新星识别方法

巡天观测与高能物理、黑洞天文等领域均有密切的联系.基于星系-超新星二分类问题,研究光谱数据预处理,结合余弦相似度改善PCA(Principal Component Analysis)光谱分解特征提取方法,用SDSS(the Sloan Digital Sky Survey)、WISeREP(the Weizmann Interactive Supernova data REPository)组成的5620条光谱数据集训练支持向量机,可以得到0.498%泛化误差的识别模型和新样本分类概率.使用Neyman-Pearson决策方法建立NPSVM(Neyman-Pearson Support Vector Machine)模型可进一步降低超新星的漏判率.     

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.     

基于图像相减和随机森林的AST3巡天暂现源及变源搜寻方法

AST3-2 (Antarctic Survey Telescopes)光学巡天望远镜位于南极大陆最高点冰穹A,其产生的大量观测数据对数据处理的效率提出了较高要求.同时南极通信不便,数据回传有诸多困难,有必要在南极本地实现自动处理AST3-2观测数据,进行变源和暂现源观测的数据处理,但是受到低功耗计算机的限制,数据的快速自动处理的实现存在诸多困难.将已有的图像相减方案同机器学习算法相结合,并利用AST3-2 2016年观测数据作为测试样本,发展一套的暂现源及变源的筛选方法成为可行的选择.该筛选方法使用图像相减法初步筛选出可能的变源,再用主成分分析法抽取候选源的特征,并选择随机森林作为机器学习分类器,在测试中对正样本的召回率达到了97%,验证了这种方法的可行性,并最终在2016年观测数据中探测出一批变星候选体.     

一种低表面亮度星系的自动搜索算法newline--YOLOX-CS

低表面亮度星系(Low Surface Brightness Galaxy, LSBG)的特征对于理解星系整体特征非常重要, 通过现代的机器学习特别是深度学习算法来搜寻扩充低表面亮度星系样本具有重要意义. LSBG因特征不明显而难以用传统方法进行自动和准确辨别, 但深度学习确具有自动找出复杂且有效特征的优势, 针对此问题提出了一种可用于在大样本巡天观测项目中搜寻LSBG的算法---YOLOX-CS (You Only Look Once version X-CS). 首先通过实验对比5种经典目标检测算法并选择较优的YOLOX算法作为基础算法, 然后结合不同注意力机制和不同优化器, 构建了YOLOX-CS的框架结构. 数据集使用的是斯隆数字化巡天(Sloan Digital Sky Survey, SDSS)中的图像, 其标签来自于$\alpha.40$-SDSS DR7 (40%中性氢苜蓿巡天与第7次数据发布的斯隆数字化巡天的交叉覆盖天区)巡天项目中的LSBG, 由于该数据集样本较少, 还采用了深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks, DCGAN)模型扩充了实验测试数据. 通过与一系列目标检测算法对比后, YOLOX-CS在扩充前后两个数据集中搜索LSBG的召回率和AP (Average Precision)值都有较好的测试结果, 其在未扩充数据集的测试集中的召回率达到97.75%, AP值达到97.83%, 在DCGAN模型扩充的数据集中, 同样测试集下进行实验的召回率达到99.10%, AP值达到98.94%, 验证了该算法在LSBG搜索中具有优秀的性能. 最后, 将该算法应用到SDSS部分测光数据上, 搜寻得到了765个LSBG候选体.     

脉冲星核成份和锥成份的宽度关系

来自脉冲星辐射束的中心和外围的成份一般分别称之为核成份和锥成份.前人研究发现脉冲星的核成份与锥成份的宽度之间有很好的0.7倍比例关系,并将此写入教科书.收集整理了33颗核、锥成份非常明显的脉冲星的观测数据,对这些脉冲星的积分脉冲轮廓进行高斯拟合和成份分离,从而得到核、锥成份的宽度及误差.分析了三峰和五峰脉冲星的核锥成份,发现核锥脉冲星其核和锥成份的宽度在统计上没有明显差别,并不存在前人发现的0.7倍的比例关系.     

Research on Pulsar Candidate Identification Method Based on Deep Residual Neural Network

As the next-generation radio astronomical telescopes continuously improve and develop, the pulsar survey will produce millions of pulsar candidates, which pose considerable challenges for pulsar identification and classification. The rapidly evolving artificial intelligence (AI) techniques are being used for pulsar identification and discovery of new pulsars. Using the pulsar data set observed with the Parkes telescope, namely the High Time Resolution Universe Survey (HTRUS), a 14-layer deep residual network has been designed (called the Residual Network, ResNet) for pulsar candidate classifications. In the HTRUS sample data, the number of non-pulsar candidates (i.e., negative samples) is much larger than that of pulsar candidates (i.e., positive samples). The imbalance between the positive and negative samples is prone to result in model misjudgement. By using the over-sampling technique to enhance the data of positive samples in the training set and adjusting the ratio of positive and negative samples, we have solved this imbalance problem. In the training process, the hyperparameters are adjusted by means of 5-fold cross validation to build the model. The test results indicate that the model can achieve a high precision (98%) and recall (100%), the F1-score can reach 99%, and that the implementation of each sample test needs only 7 ms, it has provided a feasible approach for the future big-data analysis of pulsars.     

Automated spectral classification using template matching

An automated spectral classification technique for large sky surveys is pro-posed. We firstly perform spectral line matching to determine redshift candidates for an observed spectrum, and then estimate the spectral class by measuring the similarity be-tween the observed spectrum and the shifted templates for each redshift candidate. As a byproduct of this approach, the spectral redshift can also be obtained with high accuracy. Compared with some approaches based on computerized learning methods in the liter-ature, the proposed approach needs no training, which is time-consuming and sensitive to selection of the training set. Both simulated data and observed spectra are used to test the approach; the results show that the proposed method is efficient, and it can achieve a correct classification rate as high as 92.9%, 97.9% and 98.8% for stars, galaxies and quasars, respectively.     

Pulsar candidate classification with deep convolutional neural networks

As the performance of dedicated facilities has continually improved, large numbers of pulsar candidates are being received, which makes selecting valuable pulsar signals from the candidates challenging. In this paper, we describe the design for a deep convolutional neural network(CNN) with 11 layers for classifying pulsar candidates. Compared to artificially designed features, the CNN chooses the subintegrations plot and sub-bands plot for each candidate as inputs without carrying biases. To address the imbalance problem, a data augmentation method based on synthetic minority samples is proposed according to the characteristics of pulsars. The maximum pulses of pulsar candidates were first translated to the same position, and then new samples were generated by adding up multiple subplots of pulsars. The data augmentation method is simple and effective for obtaining varied and representative samples which keep pulsar characteristics. In experiments on the HTRU 1 dataset, it is shown that this model can achieve recall of 0.962 and precision of 0.963.     

Quasar candidates selection in the Virtual Observatory era

We present a method for the photometric selection of candidate quasars in multiband surveys. The method makes use of a priori knowledge derived from a subsample of spectroscopic confirmed quasi-stellar objects (QSOs) to map the parameter space. The disentanglement of QSOs candidates and stars is performed in the colour space through the combined use of two algorithms, the probabilistic principal surfaces and the negative entropy clustering, which are for the first time used in an astronomical context. Both methods have been implemented in the voneural package on the Astrogrid Virtual Observatory platform. Even though they belong to the class of the unsupervised clustering tools, the performances of the method are optimized by using the available sample of confirmed quasars and it is therefore possible to learn from any improvement in the available 'base of knowledge'. The method has been applied and tested on both optical and optical plus near-infrared data extracted from the visible Sloan Digital Sky Survey (SDSS) and infrared United Kingdom Infrared Deep Sky Survey-Large Area Survey public data bases. In all cases, the experiments lead to high values of both efficiency and completeness, comparable if not better than the methods already known in the literature. A catalogue of optical candidate QSOs extracted from the SDSS Data Release 7 Legacy photometric data set has been produced and is publicly available at the URL http://voneural.na.infn.it/qso.html .     

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.     

On measuring the gravitational-wave background using Pulsar Timing Arrays

The long-term precise timing of Galactic millisecond pulsars holds great promise for measuring the long-period (months to years) astrophysical gravitational waves. Several gravitational-wave observational programs, called Pulsar Timing Arrays (PTA), are being pursued around the world.
Here, we develop a Bayesian algorithm for measuring the stochastic gravitational-wave background (GWB) from the PTA data. Our algorithm has several strengths: (i) it analyses the data without any loss of information; (ii) it trivially removes systematic errors of known functional form, including quadratic pulsar spin-down, annual modulations and jumps due to a change of equipment; (iii) it measures simultaneously both the amplitude and the slope of the GWB spectrum and (iv) it can deal with unevenly sampled data and coloured pulsar noise spectra. We sample the likelihood function using Markov Chain Monte Carlo simulations. We extensively test our approach on mock PTA data sets and find that the algorithm has significant benefits over currently proposed counterparts. We show the importance of characterizing all red noise components in pulsar timing noise by demonstrating that the presence of a red component would significantly hinder the detection of the GWB.
Lastly, we explore the dependence of the signal-to-noise ratio on the duration of the experiment, number of monitored pulsars and the magnitude of the pulsar timing noise. These parameter studies will help formulate observing strategies for the PTA experiments.     

Periodicity search for possible X-ray counterparts to radio-quiet gamma-ray pulsar candidates

A periodicity search of gamma-ray data is usually difficult because of the small number of detected photons. A periodicity in the timing signal at other energy bands from the counterpart to the gamma-ray source may help to establish the periodicity in the gamma-ray emission and strengthen the identification of the source in different energy bands. However, it may still be difficult to find the period directly from X-ray data because of limited exposure. We have developed a procedure, by cross-checking two X-ray data sets, to find candidate periods for X-ray sources that are possible counterparts to gamma-ray pulsar candidates. Here, we report on the results of this method obtained with all the currently available X-ray data of eight X-ray sources. Some attractive periodicity features were found. These candidate periods can serve as the target periods for a future search when new data become available, so that a blind search with a huge number of trials can be avoided.