| As the performance of the astronomical telescope improves,astronomical research has developed rapidly,and the understanding of the universe has also greatly increased.With the accumulation of data,astronomy has reached an era of "data avalanche".Large-scale sky survey projects,such as the Large Sky Area Multi-Object Fiber Spectroscopic Telescope(LAMOST)and the Sloan Sky Survey Project(SDSS),can observe huge amounts of stellar spectra.However,there are still a large number of stellar low-quality spectra that are difficult to deal with.The main characteristics of these stellar low-quality spectra include large noise,uncompleted spectrum,defective spectrum,and unclear spectral line,and so on.However,there are many valuable data such as rare objects and unknown objects in these stellar low-quality spectra,and the acquisition of them also requires a lot of manpower,time and equipment as those high-quality ones.Therefore,further processing and analysis of these stellar low-quality spectra is of great significance.Data mining can find rules and patterns that meet the conditions from a large amount of data,and machine learning can use data to optimize these patterns Therefore,data mining and machine learning have inherent advantages in processing big data,and more and more data mining and machine learning methods are applied to process and analyze massive stellar spectra.However,there is a lot of useless information,such as noise,in these stellar low-quality spectra,and the results obtained by directly using the previous methods often have a large deviation.Due to the difficulty for processing the stellar low-quality spectra,there are few methods specifically to deal with them,and the related literatures are also lacking.Therefore,how to make use of new methods to effectively process and analyze these stellar low-quality spectra is an important issue.In order to solve these problems and improve the utilization of spectra,this dissertation focuses on the data processing and analysis of stellar low-quality spectra,which include noise reduction of stellar low-quality spectra,uncompleted spectra repair,defective spectra repair,continuum fitting of low-quality spectra,rare star search from stellar low-quality spectra,and atmospheric parameter measurement of stellar low-quality spectraThe innovations and contributions of this dissertation are as follows(1)Based on the improved Generative Adversarial Network(GANs)Cycle-GANs,a deep learning algorithm Spectra-GANs is proposed.This algorithm simultaneously trains high-quality spectra and low-quality ones of homologous stars,and can effectively extract complex noise models,which can find the mapping relationship between the high-quality spectra and the corresponding low-quality ones The test results show that our method is effective in preprocessing stellar low-quality spectra.At the same time,we can also directly use the trained model to process other stellar low-quality spectra,which improves the utilization of stellar spectra.(2)According to the characteristics of stellar low-quality spectra,Monte Carlo(MC)method is introduced to fit the continuum of stellar low-quality spectra.Most of the current continuum fitting methods are used to fit stellar high-quality spectra,but not effective in fitting stellar low-quality spectra.To solve this problem,based on the Statistical Window fitting,this dissertation uses the Monte Carlo method to simulate the flux points that are filtered out,which improves the accuracy and stability of continuum fitting of stellar low-quality spectra.(3)The PCA+CFSFDP method is proposed to search for rare objects from stellar low-quality spectra.Due to the influence of noise and other factors,it is usually to search rare stars in the stellar low-quality spectra.To solve this problem,we use Principal Component Analysis(PCA)to construct a general feature spectra library to process stellar low-quality spectra.Then,a fast clustering algorithm(Clustering by Fast Search and Find of Density Peaks,CFSFDP)is used to quickly cluster these processed spectra.The results show that our method can quickly and effectively determine the candidates of rare objects,which can narrow the search range(4)A improved one-dimensional convolutional neural network(CNN)StarNet is proposed to analyze the atmospheric parameters of stellar low-quality spectra.At present,many algorithms have been successfully applied to the measurement of stellar atmospheric parameters,but it is difficult to accurately measure for stellar low-quality spectra.To solve this problem,we select the optimal number of convolutional layers and fully connected layers through experiments and expand the size of the convolution kernel to improve the StarNet.Then,the improved algorithm is used to effectively extract the high-order nonlinear features from the stellar low-quality spectra,which can improve the measurement accuracy of the atmospheric parameters of stellar low-quality spectraIn summary,this dissertation proposes methods based on data mining,machine learning,and other related technologies,which provide some new ideas and new methods to process and analyze the stellar low-quality spectra from large scale sky survey,and improve the utilization of the spectra obtained by upcoming surveys such as LSST,WEAVES and MOONS,and so on. |
