Research On Classification And Location Detection Methods Of Solar Radio Burst Events

The solar radio burst caused by the sharp increase of radio intensity due to the severe disturbance of the sun will have a huge impact on the earth's electromagnetic environment and space weather.Therefore,the study of the occurrence law,radiation mechanism and fine structure of various types of outburst events is of guiding significance for exploring the process of solar outburst activities.However,with the establishment of solar radio monitoring equipment in various countries around the world,a large amount of solar radio observation data can be accumulated,and it is difficult to quickly and efficiently classify the solar radio spectrum by manual and traditional data processing methods.This article will combine the research status of various types of solar radio burst events,and use deep learning technology to realize the identification of solar radio burst spectrum map and the automatic classification and positioning of burst types.Research on the fine structure of burst events is usually based on image processing technology,which through background denoising and data processing of the radio frequency magnitude of each frequency channel,etc.,to extract the corresponding characteristic parameters of the current burst event and study.Not only is the process cumbersome and complicated,but the parameters used in the method require manual intervention.In addition,the same processing program is not suitable for handling multiple types of burst events,and it is impossible to handle a large number of various burst events simultaneously.This thesis studies,analyzes and compares various automatic classification algorithms and target detection algorithms,and on the basis of previous methods for classifying solar radio spectrum and extracting burst feature parameters.It proposes to use a convolutional neural network with strong classification ability on the image to realize the automatic identification of the solar radio burst spectrum map,and use the target detection network to realize the automatic classification and positioning detection of various types of burst events.Finally,the position coordinates of the burst events detected by the positioning are calculated,so as to obtain its basic characteristic parameters.In the thesis,various comparative experiments were carried out on the classification of solar radio spectrum map and the location and detection of burst events.Aiming at the problem of identifying the solar radio burst spectrum map,a deep belief network and support vector machine were built to compare with the classification effect of CNN network.The results show that the classification effect of CNN method is better.In addition,according to the characteristic that the information between each frequency channel in the solar radio spectrum map is independent of each other,a 1×n convolution kernel is used for feature extraction,thereby further improving the classification performance of the network.Aiming at the problem of classification and location detection of burst types,various popular target detection algorithms are discussed,combined with the characteristics and data status of various solar radio burst events,and finally a network framework based on Faster RCNN is selected.In Faster RCNN,VGG16 and ResNet101 networks are used for feature extraction,and the network performance is compared according to the detection accuracy of the various types of burst events.In order to further improve the network's ability to detect small-scale burst events,the original Faster RCNN network is improved,and a multi-scale detection frame and multi-layer feature fusion training method are proposed.The detection accuracy of small targets is improved,so that the network in this thesis is suitable for the detection of burst events of various scales.Simple calculation of the location coordinates of the burst events detected by the positioning can get the starting and ending frequency,duration,frequency bandwidth and frequency drift rate and other characteristic parameters,which provides the possibility of real-time detection and extraction of burst events characteristic parameters.Finally,taking the type ? burst and small-scale peak burst as examples,the extraction process of basic feature parameters is explained,and a brief analysis of the burst event is made based on the statistical results of the parameters.