基于DBN,SVM和BP神经网络的光谱分类比较

发布时间：2019-06-15 17:33

【摘要】：恒星的分类对了解恒星和星系形成与演化历史具有重要的研究价值。面对大型巡天计划及由此产生的海量数据,如何迅速准确地将天体自动分类显得尤为重要。通过对SDSS DR9的恒星光谱数据进行深度置信神经网络(DBN)、神经网络和支持向量机(SVM)等算法分类的对比,分析三种自动光谱分类方法在恒星分类上的适用性。首先利用上述三种方法对K,F恒星进行识别分类,然后再分别对K1,K3和K5次型和F2,F5,F9次型识别,最后基于SVM支持向量机的二次分类模型,利用K次型的数据,构建剔除不属于K次型的模型。结果表明:深度置信网络对K,F型恒星分类效果较好,但是对K,F次型的分类效果不佳;SVM支持向量机在K,F型恒星分类以及相应的次型分类都具有较好的识别率,对K,F型分类效果要好于K,F次型的分类效果;BP神经网络对K,F型恒星以及其次型的识别一般;在剔除不属于K次型实验中,剔除率高达100%,可知SVM能够对未知的光谱数据进行筛选与分类。
[Abstract]:The classification of stars is of great value to understand the history of star and galaxy formation and evolution. In the face of large-scale sky patrol plan and the resulting massive data, how to quickly and accurately classify celestial bodies is particularly important. By comparing the star spectral data of SDSS DR9 with (DBN), neural network and (SVM) algorithm, the applicability of the three automatic spectral classification methods in star classification is analyzed. Firstly, the K, F stars are identified and classified by the above three methods, and then the K1, K3 and K5 and F2, F5 and F9 types are identified respectively. finally, based on the quadratic classification model of SVM support vector machine, the model which does not belong to the K type is constructed by using the data of the K order type. The results show that the deep confidence network has good classification effect on K, F type stars, but the classification effect on K, F type stars is not good, SVM support vector machine has good recognition rate in K, F type star classification and corresponding subtype classification, and the classification effect of K, F type classification is better than that of K, F type classification, and BP neural network is generally used for K, F type stars and the second type, and the classification effect of K, F type stars is better than that of K, F type stars, but the classification effect of K, F type stars is not good for K, F type stars, but the classification effect of K and F type stars is not good. In the K-order experiment, the removal rate is as high as 100%. It can be seen that SVM can screen and classify the unknown spectral data.
【作者单位】：三峡大学计算机与信息学院;三峡大学理学院;
【基金】：国家自然科学基金天文联合基金项目(U133113) 湖北省教育厅科研项目(B2015251) 三峡大学科学基金项目(KJ2013B018)资助
【分类号】：P144.1;TP183

【参考文献】