Research And Implementation Of High Energy Particle Classification Model Based On Improved DGCNN

The task of jet recognition in high energy physics is to identify specific signals from the background.These signals have great significance for the discovery of new particles or new processes in the Large Hadron Collider.At the same time,the technology of distinguishing different jet structures also promotes the development of quantum chromodynamics and further improves people's understanding of the field of high energy particles.In recent years,with the maturity of neural network and deep learning technology,aiming at the problems of the existing jet classification model,such as the lack of effective feature extraction means,complex training process and so on,combined with the spatial characteristics of jet data,the jet is regarded as the point cloud data in three-dimensional space.On this basis,a jet classification model based on improved dynamic graph convolution neural network(DGCNN)is proposed.The specific research work and achievements of this thesis are as follows(1)Aiming at the problems of missing and abnormal data in the opened particle classification data,this thesis firstly deals with the missing value processing and outlier detection of the dataset;secondly,combining with the professional knowledge of particle physics,the injection data features are constructed to enhance the expression ability of the data;finally,the feature selection is carried out to improve the classification accuracy.(2)In order to solve the problem of local edge information loss and spatial information loss in existing jet representation methods,a point cloud representation method in 3D space is proposed in this thesis.Point clouds have disorder and permutation invariance,which greatly restore the spatial characteristics of the jet.Therefore,the point cloud representation method can avoid the complicated data processing process,while maintaining the authenticity of the original data.(3)In this thesis,an improved DGCNN high energy particle classification model is proposed.In this thesis,K-dimensional tree(KD tree)algorithm is used instead of K-nearest neighbor(KNN)algorithm to select the nearest neighbor points,which reduces the complexity of the algorithm in the process of finding adjacent points.Then,the edge convolution module based on attention mechanism is proposed to extract local features,which improves the learning ability of the model.Finally,spatial pyramid pooling is used for multi-directional and multi-level feature sampling,in order to solve the problem caused by the different size of the input point cloud,at the same time,it can sample features from different angles according to the feature map,so that the model has better robustness.The proposed high energy particle classification model based on the improved DGCNN is compared with other jet classification models on the open particle classification dataset.The experimental results show that the proposed classification model is superior to the traditional classification model.