Research On The Application Of DBN And Its Improved Algorithm In The Classification Of Celestial Spectra

China's major Sky Survey Project LAMOST has successfully completed the eighth phase of the observation task,and obtained a large number of celestial spectral data.Confront with such a large amount of data,some traditional manual data processing methods are no longer applicable.It is an urgent need to explore and find out some high-efficiency classification methods for massive spectral data,and some classical algorithms in machine learning are more and more applicable in data classification.Using LAMOST as background,this paper will study the automatic processing and classification algorithm of the spectral data based on the spectral data released by LAMOST-DR8.The main work includes the following four parts:(1)The preprocessing and feature extraction work of spectral data.The original spectral data is processed by noise processing,interpolation,flow normalization and other pretreatment process,then the PCA and FFT are used to extract features of the data.The data is used as the input data for the subsequent experiments.(2)The application of the DBN in spectral classification.Processes the data in LAMOST-DR8 and extract the features in two ways respectively,and then the two types of data are combined in the DBN network for classification experiments.The experiment proves that the effect of combining FFT with DBN is better than that of combining PCA with DBN.(3)The application of the GA-DBN in spectral classification.Firstly,GA algorithm is used to optimize the weight value of DBN because it has strong robustness and can get the global optimum.Then,,And then use the BP algorithm to pre-train the DBN and construct the classifier.The optimized DBN is used to classify the spectrum,not only for the rough classification of spectra,but also for the sub classification of stellar spectra.After the classification experiments with different iteration and with the same iteration,it is determined that the classification effect is the best when the number of iterations is 100 and compared with different classification methods.The results show that the classification algorithm that based on the GA-DBN network has better classification accuracy.(4)The application of the SSA-DBN in spectral classification.To further optimize the DBN network structure,the SSA algorithm with stronger global optimization ability is used to optimize the weight value of DBN.The same classification experiments are carried out for the spectral data processed by GA-DBN algorithm.Compared with other algorithms,the result shows that the SSA-DBN algorithm not only has a higher classification,but also saves the running time of the experiment.It is proved that the optimization of SSA algorithm for DBN network is better than the optimization of GA algorithm for DBN network,and the experiment has achieved relatively ideal results.