Research On Solar Radio Spectrum Classification Based On Convolutional Neural Network

Solar radio is a radio radiation process with abrupt changes on the sun.It is usually divided into quiet radio,slow radio and solar radio burst.Solar radio bursts usually occur very suddenly,and usually occur simultaneously with proton bursts,X-ray bursts or solar flares in the solar active region.The radiation intensity is large and the variation is violent,which can cause a series of important geophysical phenomena,such as ionospheric disturbances,magnetic storms,etc.It affects the normal operation of modern technology systems such as spacecraft,communication and electric power.Therefore,we need to study solar radio bursts from observational data.By observing and studying the solar radio spectrum,especially the spectrum diagram of the radio burst,we can better understand the characteristics and laws of the solar radio burst,which is conducive to understanding the physics of the solar active region and the relationship between the sun and the earth,and has excellent practical significance.In the research of the classification of solar radio bursts,an improved method is proposed based on the analysis of the existing research on the classification of solar radio spectrum.The main work of this thesis is as follows:Initially,the solar radio spectrum is preprocessed.In the process of receiving solar radio spectrum data,there is a lot of noise in the spectrum image because of the interference of the receiving instrument itself and space electromagnetic wave.In addition,a large number of horizontal stripe interference and other noises are displayed in the spectral image,which will affect our ability to extract the real information of the image.The noise will be mistaken as the outbreak information,and the noise will cover up the real outbreak information,resulting in information omission.In this thesis,the channel normalization is used to remove the interference of various external environmental factors and the horizontal fringe interference caused by the channel effect of the instrument itself,at the same time,the guided filter is used to protect the outburst edge information from being blurred,and the maximum and minimum values of the image are normalized to eliminate the influence of other function transformation on the image.Furthermore,we propose two solutions to the problem of unbalanced data samples in this dataset.The first is to generate new samples and sample resampling to supplement the data categories with the small sample size.The second is to use the Weighted cross-entropy loss function to increase the contribution of small sample categories to the loss,so that the network pays more attention to the data categories with the small sample size in the training process.Finally,the good feature extraction performance of the convolutional neural network is introduced into the study of the solar radio spectrum classification.According to the characteristics of solar radio spectrum image in the experiment in this thesis,a vertical convolution kernel is proposed to replace the traditional feature extraction,so as to improve the accuracy of feature extraction of radio spectrum image.It is useful for network learning and training to classify the solar radio spectrum more accurately.The proposed method is used to test the data obtained from Huairou Observation Base of National Astronomical Observatories,Chinese Academy of Sciences.Experimental results show that the proposed method can achieve better classification results compared with other existing methods.