Deep Learning-based Special Object Search Study

A hot subdwarf star is a rare and unique celestial object with a low mass of only about half the mass of the Sun.The spectral type of hot subdwarf stars is similar to main sequence stars(O,B),but with lower luminosity and broader spectral features.Most of them are in the central helium burning phase and are located at the blue end of the horizontal branch in the Hertzsprung-Russell diagram,also known as extreme horizontal branch stars.Hot subdwarf stars are considered to be the main sources of the ultraviolet excess in the nuclei of elliptical galaxies and provide critical information for studying the interaction of binary systems and the progenitor stars of type Ⅰa supernovae.Therefore,the search for hot subdwarf stars to expand the existing sample is of great significance for the study of the structure and evolution of the Milky Way.Spectral observations of celestial bodies are one of the important means of astronomical research.Spectral data contains rich astrophysical information,such as atmospheric parameters,chemical composition,mass,and luminosity.Currently,the development of the LAMOST spectral survey project has provided a rich data foundation for the search for hot subdwarf stars.However,at the same time,the massive spectral data poses a challenge to traditional astronomical data processing methods.Machine learning techniques and methods have been successfully applied in many fields,and how to efficiently apply machine learning algorithms to process spectral data and search for hot subdwarf stars has become an important research topic.For the search of hot subdwarfs in massive astronomical spectral data,this paper focuses on the following studies:(1)A highly robust classification method for hot subdwarfs is proposed.A SeResNet+SVM binary classification model based on hybrid features is constructed for the spectral characteristics of hot subdwarfs,which can identify hot subdwarfs from other classes of objects with high accuracy,and the F1_score of the model reaches 96.17%on the test set.(2)Considering the spectral similarities of BHB-type stars,B-type stars,A-type stars,and hot subdwarfs,we further constructed a Se-ResNet+SVM four-classification model for the discriminatory results of the first step,and further filtered the hot subdwarf candidates searched by the two-classification model,and the F1_score reached 95.64%on the test set.(3)Applying the trained two-stage classification model to the 333,534 lowresolution spectra of LAMOST DR8,the binary classification model screened 3266 hot subdwarf candidates,of which 1223 were newly discovered samples.Subsequently,the four-level classifier further classified the 3266 candidates,and 409 and 296 stars were identified as newly discovered hot subdwarf candidates when the threshold values were 0.5 and 0.9,respectively.Through manual verification,we determined the numbers of truly newly discovered hot subdwarfs in these three candidate datasets to be 176,63,and 41,with accuracies of 67.94%,84.88%,and 87.60%,respectively.(4)The prediction models of the atmospheric parameters of hot subdwarfs(Teff,log g,[He/H])are constructed based on the Se-ResNet algorithm.The experimental results show that the MAE values of the model are 1212.65 K,0.32 dex,and 0.24 dex for estimating Teff,log g,and[He/H]on the test set,respectively,and then we use the model to predict the atmospheric parameter values of the newly identified 176 hot subdwarfs.In addition,the search for barium giants in the LAMOST hyperspectrum is also investigated in this paper.Barium giants are another special kind of objects in the Universe,and their study can provide a better understanding of the binary evolution theory and the mechanism of mass transfer between binary stars.In this paper,the following work is carried out for the discriminatory model of barium giants:(1)By comparing the effect of different input features on the model effect,we found that the model can have a higher F1 value when the whole spectrum is used as the feature compared to the absorption spectral line as the input feature.Therefore,we chose the whole spectrum as the input of the model.In addition,during the experiments,we also found that different classification models are needed to maximize the classification effect of barium giants for different elemental enhancements.(2)For barium giants enriched with Sr elements,we constructed a SeResNet+Stacking model for barium giant classification based on the Se-ResNet+SVM model,and the Fl score of the model was 98.07%.(3)For barium giants with Ba elemental enrichment,the best LGBM model was selected as the barium giants classification model with an F1_score of 96.87%after various classification models such as Se-ResNet+SVM,Se-ResNet+LGBM and LGBM were tested.