Modeling And Analysis Of Interplanetary Coronal Mass Ejection Automatic Detection Based On U-net

Coronal mass ejection(CME)is a frequent and violent solar activity phenomenon.When the projectiles reach interplanetary space,it is called interplanetary coronal mass ejection(ICME).ICME will interfere with the earth’s ionosphere,middle and upper atmosphere,and magnetic field,so it has a serious impact on human daily life in space exploration,satellite communication,power grid,and power facilities.At present,the identification of ICME mainly depends on manual detection,which has strong subjectivity and high time cost.Therefore,it has attracted increasing attention on how to realize data-driven ICME automatic detection quickly in the statistics and artificial intelligence community.However,it is usually complex for the frameworks of the existing ICME automatic detection algorithms.To improve the computational feasibility and detection efficiency,a new ICME automatic detection method(RU-net)is proposed by integrating multi-scale fusion and time-series segmentation together.There are skip connections between different layers of the model of RU-net,which captures the multi-scale information between different layers.RU-net is an end-to-end structure,where residual elements are embedded to accelerate the convergence of the algorithm and alleviate the problem of gradient disappearance.In addition,a simple result correction strategy is designed in the last part of the algorithm to automatically remove the impossible predicted ICME by utilizing a few ICME prior information.On the in-situ data collected by the WIND satellite,four experiments are carried out in this paper,which are the comparison experiment of missing value processing strategies,the comparison experiment of detection algorithm frameworks,the result improvement comparison experiment of the correction strategy and the result comparison experiment of feature robustness of the algorithm.The experimental results show that,the proposed RU-net has good detection performance(178 of the 230 ICMEs in the test set are successfully detected,and the F1 score is 80.18%),high detection efficiency and strong feature robustness.Finally,the statistical analysis of the experimental results reveals the characteristics of the two types of errors(FN and FP)of the predicted ICME list and verifies the physical reliability of the predicted ICME list,which is helpful to improve the method in the future.