Research On LAMOST Special Spectral Mining Algorithm Based On Deep Learning

Throughout the development of human civilization,people are constantly exploring the formation and evolution of the universe.In the universe,the spectra of celestial bodies in the optical band contains a large amount of physical information.From the spectra,a series of information such as the abundance of elements of the celestial body,movement,temperature,and distance can be obtained.In recent years,with the completion of many large sample sky survey programs at home and abroad,multiple spectral databases have been established.The release of the LAMOST(Large Sky Area Multi-Object Fiber Spectroscopy Telescope)tens of millions of spectral data sets brings new opportunities and challenges to the formation and evolution of human research celestial bodies,and also provides an opportunity for the application and promotion of deep learning algorithms.In order to better analyze and use such a huge amount of spectral data,it is necessary to establish an accurate automatic spectral classification system.So far,the research on the LAMOST classification system has been very in-depth,and many related systems have come out.However,there are still a large number of unclassified spectra labeled as "UNKNOWN" type.This paper proposes a semi-supervised spectral classification algorithm,re-sorting the "UNKNOWN" type spectra in the data released by LAMOST,and digging out a total of112,605 new spectra that can be classified,accounting for 15.65% of all "UNKNOWN" type spectra up to DR6.The main research contents and contributions of this article are as follows:1.Since the Convolutional Neural Networks performs poorly on imbalanced data sets,this paper use generative adversarial networks to enhance the data,thereby balancing the data set and improving the classification accuracy of the Convolutional Neural Networks for rare classes,and Based on this,a semi-supervised learning algorithm is proposed,and the successful implementation of this algorithm has exemplary significance for the treatment of similar problems;2.This article analyzes the "UNKNOWN" category in the data released by LAMOST,from which 294 O-type star spectra have been mined,which has more than doubled the number of original O-type star spectra;3.This article carried out an overall mining of the "UNKNOWN" type in the data released by LAMOST.A total of 112 605 spectra were identified,which greatly improved the output efficiency of LAMOST spectra.