Research On The Classification Algorithm Of Solar Radio Spectrum Based On Convolution Neural Network

At present,in the era of big data,the method of deep learning enables the knowledge behind big data to be mined and utilized.In the field of radio astronomy,Every day can produce huge amounts of astronomical observations data,how pick out effective data which researchers want is the focus of this thesis research.In this paper the method of using deep learning to deal with solar radio data selection and classification,according to the characteristics of the solar radio spectrum diagram,multi-step pretreatment of data,and put forward the two kinds of classification method based on convolution neural network,set up two kinds of solar radio spectrum classification model,realize the automatic classifying of the radio spectrum.First,We put forward a kind of based on convolution neural network solar radio spectrum automatic classification algorithm to solve laborious and subjectivity in the manual selection data,we pre-training the raw data,including data visualization,channel of normalization and down-samping,choose the suitable convolution kernels,thanks to the convolutional neural network good characteristic for feature extraction,compared with the previous work,we have a high accuracy.The experiments show that the network has a good effect on the classification of solar radio spectrum.Second,because the natural inner correlation between data set and astronomical data sets,we puts forward a combined convolution neural network and transfer learning classification method,to solve the shortage of data type and data imbalance resulting can not training deep network.we don't training the new network used limited data,only used limited data to training the new classifier,Combining the the parameters of natural image pre-training network convolution layer,pooling layer and the parameters of the newly trained classifier,form a new network.In the experiments,We've expanded the data,including random rotation,random adjustment of brightness,etc.The experimental results demonstrate the proposed method can be applied to the astronomical solar radio data,which provides the feasibility for the future study of astronomical data.