Research On Astronomical Time Series Data Classification Based On Neural Network Fusion Model CNN-G

Astronomical surveys of celestial sources produce streams of noisy time series measuring flux versus time.Unlike in many other physical domains,however,the large number of source-specific temporal gaps in data arise naturally due to intranight cadence choices as well as diurnal and seasonal constraints.In order to process these data more efficiently,it is necessary to use more accurate mining and classification techniques for noisy and irregularly sampled data with reduced manual intervention.In other applications in similar fields with huge data sets,artificial neural networks have been proven to be effective for many time series classification and prediction tasks.The usual practice is to use neural networks with fully connected layers.The assumption is that each series is considered to consist of independent observations.Among many neural networks,convolutional neural networks are very effective in speech recognition,image analysis and other fields.Their efficient optimization techniques are easier to train than other networks.At the same time,the convolutional network is very suitable for identifying local short-term patterns in time series data,but it is difficult to deal with irregular sequences of unequal length.The cyclic recurrent neural network has a unique advantage in dealing with time series of any length,but its disadvantage is also because it pays too much attention to the time effect,which causes the calculation time and calculation amount to be much higher than the convolutional neural network.Therefore,in this paper,the advantages of the two methods are combined,and the convolutional neural network and the recursive neural network are fused to form a one-dimensional convolutional neural network fusion model CNN-G,which is used to realize the classification task of astronomical time series data.The model can not only process large amounts of data,but also assist in manual operations and improve the accuracy of astronomical light source classification.The main research contents of this article are:First,collect LSST simulated astronomical time series data,and divide the astronomical light source data set into a tuple data set and a general data set accordingto the time characteristics.The tuple data set contains the basic characteristics of the astronomical light source,and the general data set contains Astronomical light source timing characteristics.Then,the exploratory data analysis is performed on the tuple data,and the differences between different light sources are found by constructing distribution histograms,box plots,cosmic coordinate maps and correlation diagrams of different types of light sources.Finally,the common data set is analyzed to construct the brightness change curves of different astronomical light sources in the six passbands and the correlation diagram between the different data columns in the two data sets to help us perform auxiliary classification of astronomical light sources.To prepare for feature selection and optimization model.Second,the traditional convolutional neural network(CNN)and recurrent recurrent neural network(RNN)are fused.Using the advantages of these two networks separately,a very long input sequence is converted into high-level features through the downsampling process,and then a shorter feature sequence that can be input into the recurrent neural network GRU layer is constructed,and a new gate-based control is constructed.One-dimensional convolutional neural network fusion model CNN-G model of cyclic unit structure.Through the feature extraction of the astronomical time series data from 2022 to 2025 simulated by the large-caliber panoramic survey telescope LSST,and constructing a one-dimensional convolutional neural network fusion model CNN-G model for training,and the traditional one-dimensional convolutional neural network CNN is compared with cyclic recurrent neural network RNN.The numerical results show that the accuracy of the model is 97.12%,compared with the CNN and RNN models improved by 2.53% and11.88%;the loss value is 0.1157,compared with the CNN and RNN models decreased by 24.82% and 34.38%;training time Compared to 21.65 min,the CNN and RNN models are shortened by 17.96% and 98.25%,respectively.It shows that the model can process massive astronomical time series data under the premise of ensuring accuracy,and has good scalability.The research in this paper shows that compared with traditional convolutional neural networks and cyclic recursive neural networks,the one-dimensional convolutional neural network fusion model CNN-G model can efficiently process astronomical time series data,effectively saving the cyclic recurrent neural network RNN in the processing sequence The time required can also make use of the time sensitivity of the recurrent recurrent neural network RNN to improve the predictionaccuracy of the classification task of simulated astronomical time series data,and further optimize the classification technology of astronomical time series data at home and abroad.