Abstract：Astronomical spectrum is an important research object in astrophysics. Many physical and chemical properties such as effective temperature, metal abundance, surface gravity and radial velocity can be inferred according to the spectra. The white dwarf main sequence binary star (WDMS) is a kind of binary star system, which is of great significance to the study of the evolution of binary stars, especially the evolution of post-common envelope. Domestic and foreign survey telescopes such as SDSS and LAMOST generate massive spectral data every day and such a large amount of spectral data cannot be analyzed manually. Therefore, it is very practical to use the machine learning method to automatically search for the WDMS spectra from the massive survey spectra. Current automatic spectral identification methods mainly depends on the existing labeled samples. Nevertheless, the number of WDMS spectra is limited. To accurately study the spectral features of WDMS spectra through a limited sample set, it is necessary to increase the number of samples and improve the accuracy of the feature extraction algorithm simultaneously. In the previous work, a batch of WDMS spectra was identified through machine learning methods in the sky survey data, providing data source for the experiment. In this paper, the generative adversarial network (GAN) is used to generate new WDMS spectra and expand the training data volume to about twice the original data set, which enhances the generalization ability of the classification model. By modifying the loss function by Anti-Bayesian learning method, the value of the loss function is correlated with the variance of the sample, which suppresses the influence of abnormally large data on the model. It improves the robustness of the model and solves the problems like vanishing gradient and getting stuck in a local optimal solution caused by the deviation of the training sample. The experiments in this paper are based on the Tensorflow deep learning library. The GAN built by Tensorflow is robust and encapsulates the internal implementation details, making the algorithm itself better represented. In addition, the Convolutional Neural Network (CNN) built by Tensor flow was used in this experiment for classification accuracy testing. The experimental results show that the two-dimensional convolutional neural network can use the convolution kernel to effectively extract the convolution characteristics of WDMS spectra and classify them. The convolutional neural network classifier based on the anti-Bayesian learning strategy achieves an accuracy of about 98.3% in the identification task of original WDMS spectra and GAN generated data. The method can also be used to search for other specific targets such as cataclysmic variable stars or supernova in the massive spectra of the telescope.
[1] Zasowski G, Johnson Jennifer A, Frinchaboy P M, et al. The Astronomical Journal, 2013, 146(4)：81.
[2] Jiang Bin, Luo Ali, Zhao Yongheng, et al. Monthly Notices of the Royal Astronomical Society, 2013，430(2): 986.
[3] Wei Peng, Luo Ali, Wang Fengfei, et al. The Astronomical Journal, 2014, 147: 101.
[4] Cui X Q, Zhao Y H, Chu Y Q, et al. Research in Astron. Astrophys, 2012, 12(9): 1197.
[5] Ren J, Luo A, Li Y, et al. AJ, 2013, 146(4): 82.
[6] WANG Wen-yu, GUO Ge-lin, JIANG Bin(王文玉，郭格霖，姜 斌，等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2016, 36(10): 3360.
[7] Thomas A, John Oommen B. Pattern Recognition, 2013, 46: 376.
[8] Goodfellow Ian J, Pouget-Abadie Jean, Mirza Mehdi, et al. Generative Adversarial Networks. arXiv: 1406. 2661.
[9] Wu Xian, Xu Kun, Hall Peter, et al. Tsinghua Science and Technology, 2017, 22(6): 660.
[10] TANG Xian-lun, DU Yi-ming, LIU Yu-wei，et al(唐贤伦，杜一铭，刘雨微，等). Acta Automatica Sinica(自动化学报), 2018, 44(5)：855.
[11] Alam S, Albareti F D, Allende Prieto C, et al. Astrophysical Journal Supplement, 2015, 219：12.
[12] Pavel Hála. Computer Science, 2014, 1412: 8341.
