白矮主序双星的光谱特征是决定其类型的关键因素, 如何有效提取恒星光谱的特征是亟待解决的问题。 提出一种新的方法, 通过构建模型捕获恒星光谱数据的特征, 对SDSS-DR10海量光谱进行自动分类。 径向基神经网络作为一种有效的计算模型, 在数值逼近和目标分类上均有较好的表现效果, 但由于目前神经网络超参数的确定大多数依赖于实验经验, 很大程度上制约了算法能力的发挥。 在分析白矮主序双星光谱数据的高维分布特征的基础上, 提出一种基于径向基神经网络的白矮主序双星自动分类模型, 并以白矮主序双星的光谱特征为导向, 针对恒星光谱提出了中心准则和宽度准则以确定神经网络的超参数, 大幅度提高了模型的准确度。 实验对分类模型进行数值训练并使用训练的模型对SDSS-DR10光谱数据进行白矮主序双星的自动分类, 共发现4 631个白矮主序双星, 通过Simbad, NED和Google交叉验证后发现其中有25个是未予以收录的新候选体。 实验结果验证了该模型在大规模白矮主序双星自动分类任务中的有效性, 新发现的白矮主序双星也为特殊天体的进一步研究补充了有效数据。

A model which is capable of capturing the spectral distribution features helps to improve the WDMS(White Dwarf + M Sequence Binaries) classification system running in SDSS-DR10 because the distribution feature of a spectra is one of the most important factors that determine its spectral type. Radial basis function (RBF) neural network is an efficient computational model that is widely used for numerical approximation and object classification. However, due to the reason that the network’s hyper-parameters are usually determined empirically, the performance of the network is limited. In this paper, on the basis of analyzing the distribution features of WDMS in a high dimensional space, an automatic classification model for WDMS ia propose based on RBF neural network. And according to the features, we propose centroids criterion and width criterion to determine hyper-parameters for the network in a more theoretical way, which improves the accuracy of the model. After training and applying the model, a total number of 4 631 WDMS candidates are classified and 25 of them are newly identified, which proves the feasibility of the model and provides further researches on WDMS with more data.