Research On Feature Extraction And Intelligent Classification Of High Dimensional Astronomical Spectral Data

With the continuous development of astronomical observation technology,more and more astronomical spectral data are collected by telescopes for astronomical research.From the spectrum,a lot of information about a celestial body can be obtained,which is meaningful for astronomers to research the formation of celestial bodies and the origin of the universe.Therefore,astronomical spectrum analysis is of great significance and value.However,in the face of the massive highdimensional astronomical spectral data,traditional manual classification methods have insufficient efficiency and accuracy.The research of the feature extraction and intelligent classification of high dimensional astronomical spectral data includ four aspects:Preprocessing method of astronomical spectrum data.First,the physical meaning of astronomical spectrum data is analyzed,after that,the data is standardized and screened under the guidance of physical meaning.Then,the SMOTE method is applied for data enhancement,which solves the problem of data imbalance.This preprocessing method helps to build high-quality data sets for training and prediction.Astronomical spectrum classification model based on autoencoder and multilayer perceptron.Aiming at the high-dimensional nonlinearity of astronomical spectral data,the autoencoder is used to reduce the dimensionality and noise of high-dimensional astronomical spectral data.In addition,a multilayer perceptron model was established to classify the encoded spectral data.The results prove that the spectral classification model established by the machine learning method can effectively classify astronomical spectra with an accuracy rate of over 79%.Convolutional neural network classification model based on residual and attention mechanism.For the local correlation characteristics of spectral data,a convolutional neural network classification model RAC-Net based on residual and attention mechanism is constructed.RAC-Net mainly includes three modules.Firstly,the local shape of the spectrum is learned by convolutional neural network;then the residual module is used to increase the depth of the network;finally,the attention mechanism is used to focus on the key channels to achieve the purpose of improving the classification accuracy.The results prove that the accuracy rate has been significantly improved and can reach more than 98%.Model lightweight improvements.For RAC-Net,the attention layer is changed to the channel weight layer.By modifying the loss function and adjusting the network structure,the original network is tailored to reduce the number of parameters to improve the efficiency of the network.Experiments show that this method can cut 40% of the parameters at the expense of 1% accuracy.