Se-ResNet+SVM model: an effective method of searching for hot subdwarfs from LAMOST 后印本

摘要: In this paper, we apply the feature-integration idea to fuse the abstract features extracted by Se-ResNet with experience features into hybrid features and input the hybrid features to the Support Vector Machine (SVM) to classify Hot subdwarfs. Based on this idea, we construct a Se-ResNet+SVM model, including a binary classification model and a four-class classification model. The four-class classification model can further screen the hot subdwarf candidates obtained by the binary classification model. The F1 values derived by the binary and the four-class classification model on the test set are 96.17% and 95.64%, respectively. Then, we use the binary classification model to classify 333,534 nonFGK type spectra in the low-resolution spectra of LAMOST DR8 and obtain a catalog of 3,266 hot subdwarf candidates, of which 1223 are newly-determined. Subsequently, the four-class classification model further filtered the 3,266 candidates, 409 and 296 are newly-determined respectively when the thresholds were set at 0.5 and 0.9. Through manual inspection, The true number of hot subdwarfs in the three newly-determined canditates are 176, 63, and 41, the corresponding precision of the classification model in the three cases are 67.94%, 84.88%, and 87.60%, respectively. Finally, we train a Se-ResNet regression model with MAE values of 1212.65 K for Teff, 0.32 dex for log g and 0.24 for [He/H], and predict the atmospheric parameters of these 176 hot subdwarf stars. This provides a certain amount of samples to help for future studies of hot subdwarfs.