Study On Stellar Spectra Classification Based On Multitask Residual Neural Network

At present,large-scale sky survey projects have accumulated massive stellar spectra.Through the analysis of the stellar spectral data,a lot of information such as type,physical parameters and position on HRD of stars can be obtained,which can help us to explore the evolution of the galaxy and even the universe.Classification of stellar spectra is the basic work of analyzing and studying stars.At present,template matching is the most popular way to classify stellar spectra.However,there are many limitations in this method.It is too dependent on existing spectral templates,and the complexity and diversity of spectral data also limit its performance.Therefore,it is of great significance to find a convenient and efficient spectral automatic processing method suitable for massive spectral data.For the sake of classifying spectra automatically,the main work of this study is as follows.(1)Experimental data preprocessing.According to the signal-to-noise ratio,stellar spectra files are filtered out to improve the data quality.Then,we interpolate and normalize the read spectral data to unify the format and range of the data.We use the Border-line Synthetic Minority Oversampling Technique(Border-line SMOTE)to solve the category imbalance problem in the dataset.In addition,different categories of spectra are labeled,and one-dimensional spectral data is folded into two-dimensional matrix,which is convenient for model operation.(2)Building multitask residual neural network.Based on the idea of multitask learning,the shared features of spectral luminosity class classification and spectral subtype classification are extracted by shared layers.In the task-specific layers,the unique features of two tasks are extracted.The loss functions of the two tasks are weighted and summed,and the optimization algorithm is used to iterate to achieve the simultaneous classification of luminosity class and spectral subtype.Structures of task-specific layers adopt the design of the residual network.The learning process of the model is transformed into a more easily optimized form through the residual mapping,which can speed up the training of the model and save the calculation cost.Appropriate evaluation metrics are selected to quantify the performance of the model,which provide the basis for the training and tuning of the model.(3)Training and testing multitask residual neural network.The adaptive adjustment of learning rate is realized by using Adam optimization algorithm,and the best value group of training epochs and batch size of the model is determined by grid search and cross validation.The classification results of the trained multitask residual neural network are verified by the test dataset.This model has achieved good performance in both luminosity class classification and subtype classification of stellar spectra.(4)Comparative analysis of experimental results.In order to further analyze the experimental results,we build and train the Random Forest(RF)and the eXtreme Gradient Boosting(XGBoost)model,and use the test dataset to verify the effectiveness of RF and XGBoost models in star spectra classification.Experimental results show that the performance of multitask residual neural network is better than these two algorithms.