Automatic Classification Method of Star Spectra Based on Semi -Supervised Mode

With the implementation of more and more large-scale spectral surveys, many star spectral data are generated, which is of great significance to the study of star evolution theory, but also brings great challenges to the traditional spectral classification and processing. In the LAMOST DR7 (v2. 0 version) spectral data set released in 2021, the total spectrum of stars has reached the order of millions, but the number of 0-type is only 129, and the data set is seriously unbalanced. Traditional machine learning classification methods cannot achieve good results for such a large amount of data and a serious imbalance of data sets. Therefore, they are mostly used to classify the spectra of adjacent stars of two types, partial types or subclasses. An automatic star spectrum classification method based on semi-supervised mode of one dimensional convolutional neural network (CNN) and one dimensional generative adversarial network (GAN) is used to solve this problem. Firstly, each spectrum is tailored and denoised, and the wavelength range of the spectrum is 370. 00 similar to 867. 16 nm. Then, uniform sampling and normalization are carried out to generate 1 X 3 700 data set samples, which are sent to one-dimensional CNN for training. In order to avoid overfitting and improve model' s prediction ability for unknown data, dropout is applied between the full connection layer and the pooling layer. This one-dimensional CNN network is used to classify six kinds of spectra except for O-type, and the average classification accuracy is 98. 08%. This experiment uses one-dimensional GAN network to solve the problem that the number of 0-type is seriously small. The input of GAN is a noise signal whose size is 1 X 900. Through the fully connected three-layer convolution operation in the generator, the output size is 1 X 3 700. The separate alternating iterative training of the generator and discriminator finally outputs the required number of 0-type star samples to expand the data set. Compared with oversampling, using data sets expanded by GAN to classify the star spectra can improve the classification accuracy of 0-type stars from 72. 92% to 97. 92%, and the accuracy of the classifier can reach 96. 28%. The experimental results show that the automatic star spectra classification method using this semi-supervised mode can realize the rapid and accurate classification of seven types of star spectra and can also be used to mine the unclassified spectra marked as "unknown".