ANN Based Spectral Classification And Stellar Atmospheric Parameter Measurement

There is abundant physical information in the celestial spectra. Through analyzing the celestial spectra, people can determine their physical and chemical composition qualitatively or quantitatively, and measure their parameters directly or indirectly. Data mining is a course of extracting cryptic, unknown potential useful information and knowledge which are embedded in the abundant, incomplete, noisy and random data. In recent years, the research of data mining algorithms of high-dimensional data has become a hot topic. Therefore, data mining technology provides a good support for celestial spectra classification and stellar atmospheric parameter measurements.With the launch of the LAMOST project, tens of thousands of spectra will be received each night. It’s difficult to deal with such massive spectra by hand. Therefore, it’s important to develop automatic processing algorithms. Compared with some traditional algorithms, neural network has a strong nonlinear mapping ability, good adaptability and fault tolerance. And it is widely used in astronomy. This paper mainly focuses on spectral classification and stellar atmospheric parameter measurement using Artificial Neural Networks (ANN).The main research jobs are as follows:(1) Feature extraction of high-dimensional spectral data. This study uses Principal Component Analysis (PCA) and Singular Value Decomposition (SVD) methods to extract the principal components from spectral data, projecting high-dimensional spectra to the low-dimensional space. This will reduce the time and space complexity of processing as well as retain the main characteristics of spectra. This paper compares the results obtained by PCA and SVD in time, extraction rate as well as the influences to classification and parameter estimation.(2) Rough classification of celestial spectra. The spectral data of SDSS are firstly projected into the different characteristic spaces, and then are classified using BP neural network model and RBF neural network model. This paper compares the time and the error rate of different models. Finally, a better method of rough classification of celestial spectra is achieved.(3) Automatic measurements of stellar atmospheric parameters. Effective temperature (Teff), gravity acceleration (log g) and chemical abundance ([Fe/H]) are three main atmospheric parameters of stars. This paper uses spectral data of ELODIE spectrum library, which are preprocessed to normalize the continuous spectrum and eliminate the singular data. Then three main atmospheric parameters are extracted from the preprocessed spectral data. This paper use BP neural network model and RBF neural network model to measure these three parameters automatically. Compare the estimated value with the real value and analyze the error. Finally, we get the most suitable parameter range where the neural network model can apply to.