Data Mining Of Celestial Spectra Based On Deep Learning

Stellar spectrum is produced by the electromagnetic wave emitted by stars through the absorption and re radiation of various gas components in the atmosphere.It contains the chemical composition,effective temperature,surface gravity and other information of stars.At present,the two telescopes with the highest spectral acquisition rate in the world are the Sloan telescope(SDSS)of the United States and the Guo Shoujing telescope(also known as LAMOST)of China.Among them,SDSS DR8 contains nearly 500000 stellar spectra,while LAMOST DR8 contains 13.28 million stellar spectra.In the face of massive spectral data,how to accurately classify them is a problem that scientists need to face.In addition,it is also a difficult problem to correctly separate rare spectra from massive spectra.Therefore,this paper attempts to explore the following methods to solve the above problems:(1)Dimensionality reduction methods of spectral data: principal components analysis(PCA),kernel principal components analysis(KPCA)and t-distributed neighborhood embedding(t-sne)are used to reduce the dimensionality of stellar spectra,After dimensionality reduction,support vector machine is used for classification.The results show that LAMOST DR5 spectral data sets m,F,K and a,G and m are linear data sets.In terms of data visualization,t-sne algorithm can achieve better dimensionality reduction visualization effect.In terms of algorithm running time,PCA takes the shortest time,KPCA takes the second place,and t-sne takes the longest time.(2)One dimensional stellar spectrum classification method: depth neural network and convolution neural network are introduced;Based on alexnet and vggnet model,a one-dimensional convolutional neural network applied to stellar spectral classification is designed.The network structure adopts 3 * 3 convolution kernel,5 convolution layers and1000 iterations.The results show that the convolution neural network designed in this paper has an accuracy of 92% in one-dimensional stellar spectrum,which is better than traditional machine learning algorithms such as support vector machine,k-nearest neighbor algorithm and Bayesian algorithm.(3)Correct separation of O-type star spectrum: In the spectrum of more than 10 million stars released by LAMOST,there are only 209 o-stars.This makes the neural network can not be fully trained due to the scarcity of data,which affects the classification accuracy of O-type star spectrum.Therefore,we expand the number of O-type stars by adding Gaussian noise.After data amplification,the recall rate of O-star classification increased by 38%.This method of data amplification can also be used in the spectra of other rare stars.